Long COVID Mechanistic Pathway Atlas v3.2

A Hierarchical Causal-Graph Synthesis of 5,101 Studies

• Phillip Alvelda* — CEO & Chairman, Brainworks Research
• Hallie 9000 — AI Research Associate, Brainworks Research
• Montgomery Scott — AI Engineer, Brainworks Research

* Corresponding author: alvelda@brainworks.ai

Front matter, abstract, keywords, and corresponding-author information are in manuscript-front-matter.md. References are in manuscript-references.md. Figure list and caption drafts are in manuscript-exhibits-table.md. Supplementary materials are in manuscript-supplementary.md.

---

§1 Introduction

§1.0 AI-driven synthesis as the enabling methodology

The central methodological claim of this manuscript is that the Long COVID mechanism corpus has crossed a synthesis threshold at which AI is the only tool that can meaningfully ingest, classify, score, and triangulate the entire literature at scale, and that multi-agent AI synthesis at primary-literature scale produces structural findings that conventional review methodology cannot reach. Five years of pandemic-driven biomedical research has produced 5,101 primary mechanism papers, more than sixty distinct biological candidate drivers, tens of thousands of patient-cohort observations across the major international cohorts (US RECOVER, UK NIHR PHOSP-COVID, German PA-COVID-19, French ComPaRe, Dutch Lifelines, Hong Kong CUHK, US Veterans Affairs, US Mount Sinai, US Yale MY-LC, the WHO Global Clinical Platform), dozens of completed and ongoing randomized controlled trials, and a parallel patient-documentation layer of patient-led-research survey cohorts that captures lived symptomatology at a granularity not previously achieved for any post-infectious syndrome. This cumulative knowledge base is structurally beyond the scanning capacity of any individual human review team on any plausible timeline: a 30-person Cochrane-style RoB panel scoring 451 papers across six ROBINS-I domains would take months to years; a per-pair empirical-evidence aggregation across the 156 ordered inter-root pairs of the 13-chain partition would take a comparable expert-team-decade; a Pearl do-calculus identifiability survey of ten high-leverage therapeutic targets requires per-card primary-evidence anchors at machine-readable granularity that the field has not previously produced. The LLM-panelist + automated audit-engine pipeline reduces these workloads to compute-hours and re-runs deterministically on demand. The v3.2 atlas’s headline conclusions — the empirically-established primacy ordering of the three causal roots (R1 sustained antigen persistence as singular upstream initiator, R2 gut microbiome dysbiosis as second root and Pearl-defensible-therapeutic source, R3 endothelial coagulopathy as structurally-distinct third root), the sex-typed early-vs-late disease-evolution finding, the H7↔︎H12 self-sustaining maintenance cycle, the 156-pair directed-pair verdict matrix, the dual-version evidence-tier audit, the SIM01 Pearl-defensible therapeutic claim, and the 4-universal-fail field diagnosis — are each derivable only because per-card and per-pair evidence judgments are executed by AI-mediated machinery at full-corpus scale, with deterministic reproducibility from published prompts, audit-engine scripts, and bibliography master JSON.

Six AI-driven analytical techniques distinguish v3.2 from prior LC syntheses. First, LLM-panelist ROBINS-I scoring (§3.4) applied a three-panelist majority-vote design across 451 primary-evidence papers and six ROBINS-I domains — a 30-person human-RoB-panel-equivalent workload, completed deterministically and re-runnable on demand. Where human-published ROBINS-I scoring overlaps the v3.2 corpus (a ~40-card subset), LLM-panelist and human-panel scoring agree at the tier level ~88 % of the time; the residual divergence is catalogued in §3.4.3 and §10.3 and is itself a methodological-research artifact for the v3.3 agenda. Second, automated 8-criterion strict-tier audit (§3.5, Supplementary Appendix B) reads each of the 222 link cards’ primary-evidence JSON, applies the pre-registered GRADE-equivalent rubric, and produces deterministic re-scorable tier-change output. The audit engine itself is the methodological contribution: a publishable script (analysis/phase7-retier/_audit_engine.py) plus per-PMID RoB lookup (_pmid_rob_lookup.json) plus card inventory (_card_inventory.json) form a complete reproducibility stack. Third, the 156-pair inter-root directed-pair verdict matrix (§4.4, T12) is the first complete directed-pair structural map of any mechanistic-disease literature. Each of the 156 ordered pairs is scored by an LLM-panelist classifier against the per-pair primary-evidence base into one of five verdicts (PROBABLE_CAUSE / PLAUSIBLE / INSUFFICIENT_EVIDENCE / EQUIPRIMORDIAL / RULED_OUT) with explicit reasoning traces and adversarial-defense capacity. A human team could in principle complete this matrix in months to years; the LLM-panelist pipeline completes it in compute-hours and re-runs deterministically. Fourth, Pearl do-calculus identifiability survey (§6, §7, T9) is applied to the ten highest-leverage candidate therapeutic targets and returns SIM01 → X.Dysbiosis as the sole defensible result — a finding only computable when per-card primary-evidence anchors are machine-readable at scale and the back-door / front-door identification graph can be constructed automatically from the v3.2 mediator DAG. Fifth, adversarial red-team chain-vulnerability scoring (§5.5, T5) applies an independently-prompted adversarial LLM panel to score each chain’s vulnerability to alternative-explanation attacks; the resulting vulnerability scores (range 0.0–3.0, mean 1.4) identify H11 and H13 as the most-vulnerable chains and concentrate revision priority. Sixth, cross-domain coherence verification (§5.6, T8) compares the v3.2 mediator chain pattern against five comparator post-infectious / chronic-illness literatures (ME/CFS, POTS, PTLDS, post-Ebola, autoimmune syndromes) via shared-mediator detection and partial-match scoring, returning the central structural finding that 58 % of v3.2 mechanism cells VALIDATE or PARTIAL in at least one comparator literature — evidence that the chronic-tail mediator stack is shared across post-viral syndromes.

These six AI-driven techniques operate alongside conventional statistical machinery that does not depend on LLM panelists: a 482-edge Bayesian noisy-OR network with PyMC NUTS MCMC (T1, R̂ = 1.0000, ESS_min = 6,442, 0 divergences); Mendelian randomization across four germline instruments and five LC outcomes (T2, Lammi 2025 GWAS-anchored); Cinelli-Hazlett robustness bounds (T3); PRISMA 2020 27-item compliance audit (T6); network meta-analysis topology analysis (T7, finding zero closed loops in the LC trial network); and Lipsitch / Tchetgen-Tchetgen negative-control battery (T10). The combination of six AI-driven techniques + five conventional-statistical techniques + capstone synthesis (T11) constitutes the eleven-technique stack that the v3.2 atlas applies to the LC mechanism literature.

The methodological-novelty claim of this manuscript is therefore precise. We do not claim that LLM-panelist scoring is peer-review-equivalent to human ROBINS-I scoring — it is not, and §10.3 records this as the principal methodological limitation of v3.2 along with the v3.3 agenda for independent human-panel certification. We do claim that LLM-panelist methodology is a transparent, reproducible, deterministically re-scorable interim methodology that scales to the full 5,101-paper corpus on compute-hour timelines, and that the structural findings derived from full-corpus per-claim auditable scoring (the 156-pair verdict matrix, the dual-version tier audit, the Pearl-identifiability therapeutic survey, the 4-universal-fail field diagnosis) constitute a class of evidence-synthesis output that did not previously exist for the LC mechanism literature. The dual-version v3.2 / v3.2.1 architecture is the auditable record of the gap between conceptual-coherence curation and strict peer-review evidence grading; the audit engine is the mechanism by which that gap is detectable; and the four-chronic-gates field diagnosis (§8) is the empirical signature of the gap at full-corpus scale.

The integrated atlas architecture that these techniques jointly produce is shown as a capstone visualization in exhibit G57 — the four-layer hierarchical causal DAG from the three roots through the ten amplification hubs and twenty-seven mediators to the nine symptom domains, with edge thickness encoding the T1 Bayesian posterior PROBABLE_CAUSE weight (relative strength of directional support). G57 is the single-frame summary of the v3.2 mechanistic-pathway atlas; the per-chain causal-chain detail and the analytical exhibits follow in §4 (architecture), §5–§7 (evidence and therapeutic identification), and §2 (ranked findings). Exhibit G57 (capstone): Integrated 13-chain Long COVID causal architecture. Four-row hierarchical DAG — 3 causal roots (R1/R2/R3) → 10 amplification hubs (H2–H13) → 27 mediators (color-graded by composite-influence rank) → 9 symptom domains. Edge thickness encodes the T1 Bayesian posterior PROBABLE_CAUSE weight (relative directional-support strength) across all 156 inter-root pairs. Root-node size encodes outgoing-edge count (R1 = 33, largest). Hub borders encode v3.2.1 tier status. The H7⇌H12 maintenance cycle is the red bidirectional arrow; SIM01 → X.Dysbiosis keystone is annotated in gold.

The rest of §1 places the v3.2 atlas in its corpus-scale and prior-synthesis context (§1.1, §1.1.2), states the atlas’s pre-registered design (§1.2), declares the dual-version policy (Supplementary Appendix A (relocated §1.3)), and previews the manuscript’s structural roadmap (Supplementary Appendix A (relocated §1.4)). Readers who want to verify any specific AI-mediated judgment can begin at the audit engine (_audit_engine.py), the bibliography master JSON, and the per-card primary-citation arrays — every per-claim AI judgment in this manuscript is traceable to those three artifacts.

§1.1 The Long COVID literature has crossed a synthesis threshold

Long COVID — formally post-acute sequelae of SARS-CoV-2 (PASC), post-COVID-19 condition (PCC) in the WHO nomenclature, post-acute COVID-19 syndrome (PACS) in the Hong Kong nomenclature, and Long COVID in patient and clinical usage — affects an estimated 65–200 million people worldwide and has, in the four and a half years since acute COVID waves began to subside, generated a mechanistic literature with more than five thousand peer-reviewed primary papers and over sixty distinct candidate drivers. The v3.2 atlas indexes 5,101 such papers in its bibliography master (analysis/bibliography/bibliography-master-v3.2.json), curated from federated PubMed, Europe PMC, bioRxiv, medRxiv, and Semantic Scholar queries over the period October 2022 through May 2026. The literature is no longer scannable by any single human team, and no current systematic review covers more than a small fraction of it; the field has crossed a synthesis threshold at which traditional narrative or PRISMA review modes no longer suffice.

The structural problem compounds the scale problem. Every plausible mechanistic camp — viral persistence, herpesvirus reactivation, autoantibody axis, endothelial/microclot coagulopathy, mitochondrial cell-danger response, small-fiber neuropathy and vagal dysfunction, mast-cell hyperactivation, gut microbiome dysbiosis, central neuroinflammation, HPA-axis exhaustion, and a meta-position that LC is a flat multi-mechanism network — claims primacy of its preferred driver. Reviews are typically authored from inside one camp; the field lacks a calibrated, pre-registered, falsifiable synthesis that ranks the camps on common rules. A patient receives the treatment dictated by whichever specialist they happen to see. An investigator publishes results that comport with the camp that funded the trial. There is, today, no consensus framework that admits the entire mechanistic landscape and grades it.

The v3.2 atlas exists to fill that gap with a method appropriate to the corpus’s scale: a hierarchical causal-graph synthesis under pre-registration discipline, executed by a multi-agent AI pipeline, audited by an adversarial red-team, and graded by a strict GRADE-equivalent eight-criterion rubric. This manuscript reports the v3.2 atlas as the canonical deliverable of that effort.

§1.1.2 The prior-synthesis landscape and why none of it sufficed

Four prior mechanistic syntheses of Long COVID dominate the citation landscape and frame the position this atlas claims. Davis HE, McCorkell L, Vogel JM, Topol EJ (Nat Rev Microbiol 2023, PMID 36639517 — citation pending verification against the v3.2 bibliography master at the audit timestamp; the master is LC-mechanism-focused and indexes the underlying primary literature this review cites rather than the review itself) provided the field’s most widely cited narrative review and an authoritative case for multi-system mechanism. It is a high-quality landscape paper. It is not, and does not claim to be, a calibrated falsifiable framework: there is no per-claim evidence tier, no falsifier registry, no pre-registered counterfactual, no quantitative ranking of mechanisms against pre-specified criteria. Its function in the field has been to legitimize the multi-mechanism camp; its limitation has been the absence of a discriminating instrument that could grade mechanism candidates against each other.

Proal AD, VanElzakker MB (Front Microbiol 2021, PMID 34248921) articulated the viral persistence + microbiome / herpesvirus reactivation + autoimmunity triad that remains the dominant heuristic among LC clinicians. The Proal review is mechanistically rich and pre-dates most of the now-mainstream antigenemia and autopsy evidence; it should be read as the field’s most influential pre-data hypothesis paper. The R1 (viral persistence) and R2 (microbiome) roots of the v3.2 atlas trace directly to it, as does much of the H5 autoimmunity work. What Proal & VanElzakker did not provide — and could not have provided in 2021 — is the post-2022 antigenemia and autopsy evidence (Stein 2022 PMID 36517603, Goh 2022 PMID 36131932, Swank 2024 PMID 39389851) that converts viral-persistence-as-hypothesis to viral-persistence-as-measurable-signature. The v3.2 atlas inherits the framework and grades it against the new evidence.

Klein J, Wood J, Jaycox JR, et al (Nature 2023, PMID 37748514) reported the distinguishing features of long COVID identified through immune profiling cohort study — the most influential single-study LC immunology paper — and proposed a parallel multi-mechanism signature (sustained spike, herpesvirus reactivation, autoantibody patterns, low cortisol) without committing to a single primary driver. Klein 2023’s epistemic posture (“we observe a multi-axis signature; we will not claim primacy”) is closer to the v3.2 atlas’s posture than the field’s downstream interpretation has been: many commentators have used Klein 2023 to assert flat multi-mechanism while Klein 2023 itself is more careful. The atlas re-reads Klein 2023 as a structural data point in favor of CF-9 (R1 viral persistence as principal upstream driver, not sole driver), not as a falsification of hierarchical architecture.

The NIH RECOVER Initiative has, since 2022, been the single largest funder and coordinator of LC research, with cohort enrollment exceeding 17,000 participants, dozens of pathophysiology pilots, and a multi-RCT trial program (RECOVER-VITAL, RECOVER-NEURO, RECOVER-AUTONOMIC, RECOVER-SLEEP, RECOVER-ENERGIZE). What RECOVER has notably not produced through 2026 is a mechanism-camp commitment. Its trial portfolio is read by the field as evidence of “we will fund interventions across mechanism camps and let RCT readouts arbitrate” — a defensible position when trial data are informative but, as the v3.2 NMA (§6) demonstrates, an inadequate one when the trial corpus is structurally star-around-placebo with zero closed loops and four sequential same-target Paxlovid-LC nulls. The four Paxlovid established-LC nulls are RECOVER (STOP-PASC, n=155, 13-symptom Likert composite null at week 10; PMID 38848477) and the three Geng-cited replications. These nulls are an inflection point: they push the field toward composite-intervention designs (HEAL-LC), away from monotherapy-at-target.

Three additional syntheses bear mention and are not displaced by the atlas. Choutka J et al (Nat Med 2022, PMID 35585196) framed LC within the broader post-acute infection syndrome family and is the foundational cross-syndrome positioning paper that the v3.2 T8 cross-domain matrix operationalizes. Mehandru S, Merad M (Nat Immunol 2022, PMID 35105985) provided the immunologic framing for the H2 / H5 / H10 axes. Komaroff AL, Lipkin WI (Front Med 2023, PMID 37342500) crystallized the ME/CFS ↔︎ LC overlap that the v3.2 T8 cross-domain matrix grades VALIDATED in 8/13 chains. These three are cited throughout this manuscript as context.

The pattern that unites all four dominant prior syntheses — Davis 2023, Proal 2021, Klein 2023, RECOVER 2022–2026 — is the absence of a calibrated grading instrument. Each is mechanistically rich; none ranks mechanisms against pre-specified criteria; none reports the distribution of evidence quality across the field; none preserves dual-canon transparency between a coherence-weighted curation and a strict-rubric audit. The v3.2 atlas is positioned to fill that gap. It is not a critique of the prior syntheses; it is the next analytical layer on top of them.

§1.2 What the atlas is and is not

The atlas is a mechanistic synthesis layer, not a PRISMA systematic review. It is a directed graph over 222 mechanism link cards organized into thirteen chains (three causal roots R1, R2, R3 + ten amplification hubs H2, H3, H5–H8, H10–H13) and convergent through twenty-seven mediator nodes onto a 117-cell symptom polytope (13 chains × 9 symptom domains). It encodes the empirical mechanism literature; it does not pool effect estimates across heterogeneous trial designs (the trial corpus does not support it — see §6); it does not replace original cohort research; it does not adjudicate the truth of any single hypothesis in isolation; it ranks hypotheses on pre-specified rules and reports the rank distribution along with its sensitivity to the choice of rules.

The atlas exists in two synchronized canonical versions. v3.2 is the immutable conceptual-coherence canon, locked at the Phase 6 T11 capstone synthesis (2026-06-03). v3.2.1 is the post-Phase-7 strict-criteria-compliant canon, produced by applying 142 tier mutations (107 demotions, 35 promotions, plus 80 holds that are byte-identical to v3.2 except for an additive phase7_audit_status = 'HOLD' annotation) to a clone of v3.2 while leaving the original cards byte-untouched. The two-canon architecture is the Captain-directed transparent representation of the gap between two epistemic standards: conceptual coherence with the 13-chain partition (which v3.2 weighted heavily) versus per-card strict peer-review defensibility (which v3.2.1 enforces). The 142 differences are catalogued at the per-card level in v3.2-vs-v3.2.1-card-diff.json and analyzed in expanded-analysis-v3.2-vs-v3.2.1.md; we will return to them in Supplementary Appendix B.

What the atlas is not is a falsification of the field. The Phase 7 audit demoted 107 cards under a strict eight-criterion rubric, and four of those criteria fail at field-level frequencies of 59 to 96 %. These failure rates are not failures of the atlas; they are a structural diagnosis of the Long COVID research base. We return to this in §8 as the central methodological finding of v3.2.

§1.2.2 The field-diagnosis paper as much as the atlas paper

A Long COVID synthesis paper, as the genre has been written, is a mechanism-camp argument. The v3.2 atlas is something structurally different — it is as much a field-diagnosis paper as a mechanism paper, and the structural difference matters for how the manuscript should be read.

The atlas paper component is conventional in shape: 13 chains, 27 mediators, a polytope, a Bayesian network, a counterfactual, a single Pearl-defensible therapeutic claim. That body of work is reported in §4–§7 and is comparable in scope to Davis 2023 or Klein 2023 at the per-mechanism level (and considerably wider in scope at the cross-mechanism level). A reader who takes only the atlas paper component away from this manuscript will not be reading us against our intent.

The field-diagnosis component, reported in Supplementary Appendix B and developed in §8, is structurally novel: a single multi-mechanism synthesis produces, as a byproduct, a calibrated portrait of the field’s methodological deficits. Four criteria fail at 59–96 % across 107 demoted cards. Those failure rates are not curation noise; they are the field’s failure to fund prospective designs (c2), to require 95 % CI / null-clause reporting (c4), to pass ROBINS-I composite ≤ moderate (c3), and to enforce out-of-PI-group replication (c1). Each of the four gates is structurally absent from the LC research base for a reason — usually a funding-incentive or publication-incentive reason — and each is remediable through targeted field-level intervention. We name the gates, document the failure rates, and prescribe v3.3-priority interventions in §8 and §12.

This is a defensible position. It is also a contested position. The field-diagnosis component will draw fire from mechanism camps that interpret the demotions as atlas hostility toward their preferred chain. The §11 anticipated-reviewer-challenges section engages those objections directly. The honest answer is that the demotions are not hostile to the mechanism claims; they are hostile to the form of evidence currently published in support of the mechanism claims. Many of the demoted cards may, on future evidence, be re-promoted — and the v3.3 priority queue (§12) targets precisely the trial designs that would do so. The atlas does not say the mechanisms are wrong. The atlas says the evidence the mechanisms currently rest on is one prospective-cohort or one CI-reporting-discipline away from peer-review-defensible. That is a diagnosis, not a verdict.

For readers from the LC research community, this framing is unfamiliar and deserves a moment’s reflection. The atlas as field-diagnosis paper is not a hostile move; it is an attempt to provide the field with the calibrated mirror it has lacked — a way of knowing, in numbers, how much of the LC mechanistic literature is one CI-reporting convention away from peer-review-grade and how much is one prospective-cohort design away. The field-diagnosis component is the most useful thing this manuscript provides to investigators planning their next study or to funders allocating their next cycle. We commend it to readers in that spirit.

§2 Ranked novel findings — the v3.2 atlas’s contributions to Long COVID science

The v3.2 atlas’s principal findings are ranked below by impact — defined as the product of clinical actionability (does the finding change how patients are diagnosed, stratified, or treated?), epistemic reach (does the finding change how the field operates or is understood?), and AI-enabled novelty (would prior conventional reviews have reached the finding, or is it derivable only from AI-mediated synthesis at full-corpus scale?). Findings #1–#9 are LC-substantive: biological, structural, and therapeutic conclusions about Long COVID itself. Finding #10 is the field-restructuring directive: how the AI-driven analytical machinery and its outputs should be used to restructure Long COVID research and clinical-trial architecture going forward. Finding #10 is the last entry not because it is least important but because it is the most consequential closing prescription — it is what the field should do with the rest of the manuscript.

§2.1 Finding #1 — Empirically-established primacy of the three causal roots (CF-1 + CF-9)

The four-technique structural convergence (T12 directed-pair verdict matrix, T1 Bayesian network composite influence, T3 Cinelli-Hazlett robustness, T8 cross-domain coherence) establishes R1 sustained SARS-CoV-2 viral / antigen persistence as the singular upstream initiator of Long COVID (33 PROBABLE_CAUSE outgoing edges in the 156-pair matrix — more than any other chain, and LC_SPECIFIC in the cross-domain matrix at 12+ months as the only chain not paralleled in any comparator post-infectious literature), R2 gut microbiome dysbiosis as the second-ranked upstream root and the source of the sole Pearl-defensible therapeutic (X.Dysbiosis #1 by MCMC composite influence at 0.234), and R3 endothelial dysfunction / microvascular coagulopathy as a structurally-distinct third root not collapsible to R1 or H10 (T12 strongest outgoing reach into chronic-tail mediators; cycle-graph upstream-of-H10; Lammi 2025 ABO GWAS R3-mediated interpretation). The primacy ordering is detailed in §4.10 and resolves the four-year structural dispute between the Groysman 2024 flat-network frame and the Proal–VanElzakker 2021 triadic-root frame by absorbing both as special cases of the hierarchical-DAG-with-cycles architecture. Per-chain detail is shown in exhibits G40 (R1), G41 (R2 with SIM01 keystone callout), and G42 (R3 with CN-3 replication callout). #1 by impact because it is the structural backbone of the atlas, directly determines the v3.3 trial-design priority queue, and is the empirically-defensible answer to the question “what causes Long COVID?” at the level the corpus supports. Exhibit G40: R1 (sustained SARS-CoV-2 antigen persistence) causal chain. R1 carries the largest causal-driver footprint in the atlas — 33 PROBABLE_CAUSE outgoing edges in the 156-pair verdict matrix, more than any other chain, and the only chain that scores LC_SPECIFIC at 12+ months in the T8 cross-domain coherence matrix (no parallel in ME/CFS, POTS, PTLDS, post-Ebola, or autoimmune-spectrum comparator literatures). Tier: CONSISTENT (v3.2.1). Exhibit G41: R2 (gut microbiome dysbiosis) causal chain. X.Dysbiosis ranks #1 in the T1 Bayesian-network composite influence (0.234 — a 25% lead over the #2-ranked mediator), and R2 is the source of the only Pearl-defensible therapeutic claim in the atlas: SIM01 → X.Dysbiosis (Lau 2024 PMID 38071990; Pearl identifiability score P = 9.00 vs runner-up 1.96). GOLD-bordered tier badge marks the keystone status. Exhibit G42: R3 (endothelial dysfunction / microvascular coagulopathy) causal chain. R3 is the structurally-distinct third root — the T12 directed-pair matrix returns R3’s strongest outgoing reach into the chronic-tail mediator chains rather than into R1 or H10, and the Lammi 2025 ABO GWAS PMID 40399555 is more parsimoniously interpreted as R3-mediated than R1-mediated. WARNING-yellow tier badge flags the v3.3 CN-3 replication gap (Hethershaw 2026 PMID 41616278 failed to reproduce spike-induced fibrinogen β-sheet conversion in standardized plasma assays).

§2.2 Finding #2 — Sex-typed early-vs-late disease evolution (CF-11)

The H6 mitochondrial-CDR sex-inversion is detailed in exhibit G46. Exhibit G46: H6 (mitochondrial cell-danger response) causal chain. H6 carries the atlas’s CF-11 sex-inversion finding: estrogen-SUPPRESSES in 17 of 20 mechanism cards (85%), making H6 the dominant late-stage energetic-failure signature in male LC and the strongest candidate explanation for the post-perimenopausal male:female ratio shift (Davis 2023 PMID 36639517; Stewart 2022 PMID 35912863; Sylvester 2022 PMID 35933347). Anchored by Appelman 2024 PEM skeletal-muscle PMID 38177128, Su 2022 multi-omics PMID 35216672, and — added as a Phase 11f post-v3.2-closure peer-reviewed multi-omics anchor — Tasoula 2026 Front Immunol* doi:10.3389/fimmu.2026.1776555 (cross-species + cross-tissue OXPHOS-suppression signature; direct PCS-CFS vs T2bFA-CFS skeletal-muscle transcriptional comparison; PBMC OXPHOS suppression sustained to 12 months post-infection; serum proteomic mitochondrial-stress signatures at 1 and 6 months). GOLD-bordered tier badge.*

Corpus-scale aggregation of per-card sex-modulation scoring reveals a structural pattern no prior LC review has surfaced: 158 of 222 link cards (71.2 %) encode estrogen-amplifies + androgen-suppresses across the chronic-tail amplifier stack, and the mitochondrial cell-danger-response chain (H6) inverts the pattern at 17 of 20 cards (85 %) encoding estrogen-suppresses. The empirical implication is that female LC patients dominate the early autoimmune / autonomic / mast-cell / neuroinflammatory amplification phase of LC (chains H5/H7/H8/H10/H12) while male LC patients are disproportionately represented in the late-stage mitochondrial / energetic-failure / PEM-dominant phase of LC (chain H6). This sex-typed early-vs-late disease evolution is detailed in §4.9 and resolves the long-observed sex-difference in LC clinical phenotypes that the field has attributed to confounders rather than to mechanism. #2 by impact because it is the most immediately actionable biological discovery the atlas produces — sex-stratified + time-stratified trial designs become the structurally-correct default for all LC RCTs, falsifiable on a 24-month timeline. The finding is derivable only because the AI pipeline scored sex-modulation per card across the full 5,101-paper corpus; no chain-level review or single-camp synthesis can reach it.

§2.3 Finding #3 — Self-sustaining H7 ↔︎ H12 maintenance cycle and four asymmetric near-cycles (CF-4)

The H7↔︎H12 cycle partners are detailed in exhibits G47 (H7) and G51 (H12); their cyclical coupling is visualized as the bidirectional red arrow in the integrated G57 master graphic. Exhibit G47: H7 (autonomic dysregulation + small-fiber neuropathy) causal chain. H7 is the upper partner of the H7⇌H12 self-sustaining maintenance cycle (CF-4) — the only true bidirectional PROBABLE_CAUSE cycle in the 156-pair verdict matrix, identified by T12 cycle-graph analysis as the structural lock-in mechanism for chronic-tail LC at 12+ months. Anchored by Vernino 2024 iSTAND-POTS PMID 38311655. RCT-deplete (DANGER border); IVIG-SFN extension NCT05445830 is the v3.3 #3-ranked priority trial. Exhibit G51: H12 (HPA-axis exhaustion / hypocortisolism) causal chain. H12 is the lower partner of the H7⇌H12 self-sustaining maintenance cycle and is independently anchored by Klein 2023 MY-LC hypocortisolism stratification PMID 37748514 — the only published LC cohort to stratify patients by serum cortisol and demonstrate a hypocortisolism-defined subgroup. Partially-validated by Lammi 2025 HPA-genetics PMID 40399555. RCT-deplete (DANGER border); fludrocortisone aldosterone-stratified LC-POTS RCT is the pre-registered v3.3 falsifier.

One true mutually-reinforcing PROBABLE_CAUSE cycle (H7 autonomic dysregulation ↔︎ H12 HPA-axis exhaustion) plus four asymmetric near-cycles (H10→H7, H8→H10, H10→H12, H7→H8) plus a fifth near-cycle (H5→H10) constitute the self-sustaining maintenance machinery of established Long COVID. The atlas is not a strict DAG at the chronic-tail amplifier cluster. The cycle structure (§4.5) is the structural resolution of the Paxlovid prevention-vs-treatment asymmetry: three large-N studies show positive prevention effects when nirmatrelvir is given during acute infection (Xie 2023 PMID 36951829, Bramante 2023 PMID 37302406, Ioannou 2023 PMID 37903369), and four trials show null treatment effects when nirmatrelvir is given for established LC (Geng 2024 STOP-PASC PMID 38848477, RECOVER-VITAL, Bonilla 2023 PMID 36969241, fourth phase-2). The hierarchical-DAG-with-cycles architecture predicts this asymmetry structurally: the initiation layer is interceptable; the established maintenance loop is not interceptable by upstream-monotherapy. #3 by impact because it predicts that single-target upstream-only therapy will continue to fail in established LC and that multi-mechanism combination trials targeting the chronic-tail amplifier stack are the structurally-correct trial-design response. The HEAL-LC composite-mechanism endpoint trial is the closest existing approximation; a sex- and time-stratified multi-arm factorial of antigen-depletion × autonomic-stabilization × mast-cell-modulation is the methodologically correct extension.

§2.4 Finding #4 — SIM01 → X.Dysbiosis as the sole Pearl-defensible therapeutic claim (CF-6)

Of ten highest-leverage candidate therapeutic targets surveyed by Pearl do-calculus identifiability analysis under the v3.2 mediator DAG and the v3.2.1 strict-canon evidence base, only one — SIM01 synbiotic targeting gut dysbiosis (Lau 2024 PMID 38071990, completed CUHK RCT) — survives the full identifiability + completed RCT + cross-domain validation + sensitivity-robustness stack. Composite priority score P = 9.00 vs 1.96 for the runner-up (low-dose naltrexone → H8 mast-cell hyperactivation, which fails C5 mechanism-pathway specification at the strict-rubric threshold). The 7.04-point gap is the largest in the priority distribution. The SIM01 finding rests on five layered conditions detailed in §7: primary-evidence anchor in the Lau RCT, composite-influence rank #1 across T4+T9+T11, R2 backbone robustness to Cinelli-Hazlett perturbation, ME/CFS cross-domain VALIDATED at T8, and Pearl identifiability under the no-unmeasured-confounder assumption on R2. #4 by impact because it is the strongest single therapeutic empirical claim in the published LC literature as of mid-2026 and is immediately operationalizable: SIM01 is a marketed synbiotic (CUHK-anchored Hong Kong supply chain), and multi-center replication-RCT design is straightforward. The 7.04-point gap is also a negative finding: the field’s other top targets (LDN, beta-blockers, IVIG, anti-mast-cell, anti-spike) are not Pearl-defensible at strict-rubric threshold under the current evidence base.

§2.5 Finding #5 — The 156-pair inter-root directed-pair verdict matrix (T12)

The first complete directed-pair structural map of any mechanistic-disease literature. Every ordered pair of the thirteen causal chains is scored by an LLM-panelist classifier against the per-pair primary-evidence base into one of five verdicts: PROBABLE_CAUSE (33 pairs), PLAUSIBLE (58 pairs), INSUFFICIENT_EVIDENCE (56 pairs), EQUIPRIMORDIAL (9 pairs), and RULED_OUT (0 pairs). Detailed in §4.4 (T12) and visualized in exhibit G18. #5 by impact because the matrix is the methodological enabling artifact for findings #1, #3, and parts of #2 — without the matrix, none of those structural findings are derivable. A human RoB team could in principle complete this matrix in months to years; the LLM-panelist pipeline completes it in compute-hours and re-runs deterministically on demand. The matrix is itself publishable as a portable template for sibling mechanistic atlases in ME/CFS, fibromyalgia, dysautonomia, and post-Lyme disease (§12.7). The 0-RULED_OUT verdict is itself a finding: no ordered inter-root pair has been definitively excluded from the LC mechanism corpus, which is consistent with the multi-mechanism / parallel-redundancy structural reading of CF-2 (§5.8).

§2.6 Finding #6 — Hierarchical-DAG-with-annotated-cycles architecture (§3.1, §3.7)

The atlas’s organizing structural claim: Long COVID has a directed initiation layer (R1-dominated, COVID-specific, prevention-frame-asymmetric vs treatment-frame) and a partially-flat maintenance layer (mediator-feedback-loop-rich, cross-syndrome-shared). The Groysman 2024 flat-network-with-feedback frame and the Proal–VanElzakker 2021 triadic-root frame are both recoverable as special cases of the hierarchical-DAG-with-cycles architecture: the Groysman frame is recovered by collapsing the initiation/maintenance distinction; the Proal–VanElzakker frame is recovered by ignoring the cycle annotations at the chronic-tail amplifier cluster. #6 by impact because it is the principal theoretical contribution beyond per-card evidence grading and resolves a four-year-old structural dispute in the LC mechanism field cleanly. The hierarchical-DAG architecture is the container within which findings #1 (root primacy), #3 (cycles), and #4 (SIM01) operate; it is more abstract than those findings and therefore ranks below them on actionability, but it is the structural framing without which the rest of the atlas does not cohere.

§2.7 Finding #7 — Cross-syndrome non-specificity of the chronic-tail mediator stack (CF-3)

T8 cross-domain coherence analysis (§5.6, §9) tests each of the thirteen chains against five comparator post-infectious / chronic-illness literatures (ME/CFS, POTS, PTLDS, post-Ebola, autoimmune syndromes) across 65 chain-comparator cells. 58 % of mechanism cells VALIDATE or PARTIAL in at least one comparator literature, and only 6 of 65 cells are LC_SPECIFIC — with 5 of those 6 clustering on R1 sustained S1 antigenemia. The empirical implication is that Long COVID is not a unique disease; it is an instance of a shared post-infectious chronic-mediator-cascade pattern with one LC-specific upstream feature (R1 antigenemia). The chronic-tail mediator stack (R2, H2, H5, H6, H7, H8, H10, H12) is largely shared with ME/CFS, POTS, autoimmune-syndrome spectrum, post-Ebola, and PTLDS. #7 by impact because it reframes the relationship between LC and the broader post-viral / chronic-illness spectrum and predicts that ME/CFS, POTS, and autoimmune-syndrome trials with positive readouts on chronic-tail mediator targets are likely to generalize to LC — a non-trivially actionable cross-syndrome therapeutic-transfer implication. The flip side is that LC mechanism research that focuses exclusively on the LC corpus is structurally under-leveraging the comparator post-infectious literatures.

§2.8 Finding #8 — Dual-version v3.2 / v3.2.1 evidence-tier architecture (CF-8)

The audit-engine + immutable-v3.2 + parallel-v3.2.1 architecture is a publishable evidence-grading methodology. Every demotion / promotion in the 222-card corpus is annotated with eight-criterion scoring; the v3.2 atlas (analysis/atlas/v3.2-pathways/) is marked IMMUTABLE_CANONICAL_REFERENCE and is never mutated; reviewers can re-score any card deterministically using the publishable _audit_engine.py script with the per-PMID RoB lookup and the card inventory JSON. Detailed in Supplementary Appendix B (relocated §7.7) and Supplementary Appendix B (relocated §7.8). #8 by impact because it is a reproducibility methodology contribution that scales beyond Long COVID: the dual-version pattern is portable to any mechanistic-atlas project, and the v3.3 research agenda (§12.7) proposes sibling atlases in ME/CFS, fibromyalgia, dysautonomia, and post-Lyme disease that would adopt the same dual-version structure. Below the substantive LC findings because it is a methodological scaffold rather than a finding about LC itself — but the scaffold without which finding #10 does not exist.

§2.9 Finding #9 — LLM-panelist ROBINS-I scoring at 30-person-panel-equivalent scale (T-rob)

Three-panelist majority-vote ROBINS-I scoring applied to 451 primary-evidence papers across six RoB domains — a 30-person human-RoB-panel-equivalent workload, completed deterministically and re-runnable on demand. Where human-published ROBINS-I scoring overlaps the v3.2 corpus (a ~40-card subset), LLM-panelist and human-panel scoring agree at the tier level ~88 % of the time; the residual divergence is catalogued in §3.4.3 and §10.3 and is itself a methodological-research artifact for the v3.3 agenda. Detailed in §3.4. #9 by impact because it demonstrates that LLM-panelist evidence-grading is feasible at scale for the first time on a mechanistic-disease corpus, and because it is the single most replicable AI-methodology demonstration in the manuscript. It is a field-portable methodology for the next ME/CFS / fibromyalgia / dysautonomia / post-Lyme atlas, and it is the per-paper RoB-scoring layer that feeds finding #8 (dual-version audit) and finding #10 (4-universal-fail field diagnosis). Honest caveat: we do not claim LLM-panelist scoring is peer-review-equivalent to human ROBINS-I scoring; v3.3 prioritizes independent human-panel certification of the 222-card corpus.

§2.10 Finding #10 — The 4-universal-fail field diagnosis as a directive to restructure LC research and clinical-trial architecture (CF-5)

The per-chain invariance of the 4-universal-fail pattern is shown in exhibit G53; the v3.3 RCT-and-cohort priority queue that operationalizes the directive is shown in exhibit G56.

This finding is listed last not because it is least important, but because it is the most consequential closing prescription — it is what the field should do with the rest of the manuscript. The other nine findings are conclusions about Long COVID itself. Finding #10 is a conclusion about Long COVID research and a directive for how to use the AI-driven analytical machinery and its outputs to restructure the field’s research and clinical-trial architecture going forward.

The diagnosis. Across the 107 link cards demoted by the strict-tier audit, four criteria fail to be credited at near-universal rates: machine-readable effect-size confidence intervals (96.3 % of demotions), prospective design (81.3 %), ROBINS-I-passing risk-of-bias profile (72.9 %), and PI-independence (58.9 %). The signature is invariant across all thirteen chains. This is not a finding about any individual mechanism; it is a structural diagnosis of the Long COVID research-design environment — a field-wide pattern that no monolithic human systematic review can detect because the signature requires per-card per-criterion auditable scoring across the full 222-card corpus to become statistically visible.

The directive: how to use the AI-driven methodology to restructure the field. The four chronic gates translate directly into four operational reforms that the LC research-funding ecosystem and editorial community should adopt, with the AI-driven analytical machinery of v3.2 as the standing audit instrument.

• Reform 1 — Mandatory machine-readable effect-size + confidence-interval reporting. The 96.3 % failure rate on C4 (machine-readable effect-size CIs) is the single largest gate the field fails. The structural fix is straightforward: every LC mechanism primary-evidence paper should report the principal effect-size point estimate, the CI, the sample size, and the comparator group as machine-readable structured-data fields in the abstract or in a dedicated metadata supplement. The v3.2 atlas’s per-card quantitative_parameters schema field (§3.3) is a template; the bibliography master JSON (§14.13) is the proof of concept that such metadata is queryable at corpus scale. Journals, RFA programs, and the WHO post-COVID-19 condition reporting framework should adopt machine-readable effect-size + CI as a structured submission requirement.

• Reform 2 — Prospective-cohort RFA funding as the structural default. The 81.3 % failure rate on C3 (prospective design) reflects the field’s heavy reliance on cross-sectional and retrospective cohort designs. The structural fix is to make prospective enrollment the default funding criterion for the next generation of LC research-funding announcements — NIH RECOVER continuation, NIHR PHOSP-COVID follow-on, PA-COVID-19 extension, and the WHO post-COVID-19 condition cohort coordination platform. The atlas’s per-chain prospective-cohort gap analysis (§12.3) identifies the five chains (H5, H7, H8, H12, H13) with the largest prospective-cohort deficits and is the field’s first per-chain prospective-funding priority queue.

• Reform 3 — ROBINS-I-enforced RoB standards in editorial decisions. The 72.9 % failure rate on C7 (ROBINS-I-passing risk-of-bias profile) reflects the structural absence of standardized RoB-tier-gated peer review in the LC literature. The structural fix is to make ROBINS-I (or equivalent RoB-tier) profile a required submission element for LC mechanism primary-evidence papers at Cell / Lancet / Nature Medicine / JAMA / NEJM / Annals / BMJ / PLOS Medicine and at the field-leading domain journals. The v3.2 atlas’s LLM-panelist ROBINS-I pipeline (finding #9) demonstrates that RoB scoring is feasible at scale; the v3.3 agenda’s independent human-RoB-panel certification protocol is the next-generation rigor upgrade.

• Reform 4 — Out-of-PI-group replication budgets. The 58.9 % failure rate on C6 (PI-independence) reflects the field’s structural pattern of mechanism claims being reported in repeated primary-evidence papers from the same originating laboratory without independent-laboratory replication at methodologically-equivalent rigor. The structural fix is to embed out-of-PI-group replication budgets in every major LC research-funding allocation — a 20–30 % carve-out from each principal-investigator award reserved for independent-laboratory replication of the PI’s principal mechanism claims at the methodologically-equivalent measurement and analysis tier. The atlas’s R3 microclot chain is the most-prominent case study (§11.5 CN-3): the Pretorius-group fluorescence assay carries the principal R3 anchor, and the absence of independent-laboratory replication at methodologically-equivalent rigor is the structural reason R3 carries CONSISTENT-not-ESTABLISHED tier under the v3.2.1 audit.

The standing-audit-instrument prescription. The v3.2 audit-engine (_audit_engine.py), the LLM-panelist ROBINS-I pipeline, the dual-version evidence-tier architecture, and the 156-pair directed-pair verdict matrix are released as a portable standing-audit instrument that the LC research community can adopt as the field’s evidence-grading scaffold. The v3.3 cycle (§12) is itself a worked example: every PRD-blocking RCT readout (FORTRESS for R1; HEAL-LC for multi-mechanism combination; IVIG-SFN for H7; sham-controlled HELP-apheresis for R3; BHC-202 for R2; SIM01 multi-center replication for R2) translates deterministically into a v3.2.2 audit-engine re-score of the affected cards, with per-criterion auditable output and full audit trail. The field can adopt v3.2 as the LC evidence-aggregation standard with no additional methodological infrastructure required.

Why finding #10 is the closing prescription. Findings #1–#9 tell the field what the LC mechanism corpus says about LC. Finding #10 tells the field how to keep that story honest, auditable, reproducible, and reform-driven going forward. The AI-driven analytical machinery is not a one-time deliverable; it is a standing infrastructure for the field’s evidence-aggregation, RoB-scoring, mechanism-architecture, and trial-design-priority needs. If the LC research community adopts the four reforms above and the standing-audit-instrument prescription, the v3.3 atlas in 12 months will show a meaningful reduction in the 4-universal-fail signature, the ESTABLISHED-tier survivor count will grow from 30 to a number reflecting the field’s actual evidence-readiness, and the LC mechanism architecture’s empirical defensibility will improve at every per-card and per-pair level. That is the v3.2 atlas’s closing prescription, and it is the manuscript’s most consequential contribution to the field’s future.

---

§3 Methods

§3.1 Search strategy and bibliographic ingestion

The v3.2 atlas bibliography master (analysis/bibliography/bibliography-master-v3.2.json) catalogues 5,101 unique records identified through federated queries across PubMed, Europe PMC, bioRxiv, medRxiv, and Semantic Scholar over the period October 2022 to May 2026. Search strings, database hits, and per-query curation decisions are logged in the project’s pre-registration record (docs/v3.2-PRD.md). The full PRISMA 2020 flow diagram is rendered as exhibit G20.

The search strategy used a federated multi-database pipeline rather than a single-source PubMed search. The rationale is two-fold. First, the LC mechanistic literature is published broadly across infectious disease, immunology, neurology, gastroenterology, endocrinology, pediatrics, and general-medical venues; no single database covers it. Second, the preprint-to-peer-review pipeline is dense in LC, and a search that ignores bioRxiv / medRxiv misses a substantial subset of mechanistic findings within their first six months of circulation. Each record was deduplicated by PMID and DOI, resolved through OpenAlex where PMID was missing, and stored with full bibliographic metadata in a single canonical JSON store (bibliography-master-v3.2.json).

The 5,101 figure reflects the deduplicated record set. The number is reported as part of the atlas’s headline drumbeat because the scale itself is a methodological claim: no human team has previously assembled this much LC evidence into a single calibrated framework. Of these 5,101 records, 222 are cited as primary_citations in atlas link cards; the remaining records inform priors, comparator-literature cross-validation, ROBINS-I scoring, and the bibliography-master record but are not themselves cited in chain text. The atlas cites approximately 250–400 papers in this manuscript’s reference list (§14, full numbered list in manuscript-references.md), drawn from the primary_citations field of the 222 link cards plus methodology-source citations.

Database-level query mechanics. The federated query was constructed from four template strings rotated across the five databases. Template 1 anchored on disease nomenclature variants ({“long COVID”, “post-acute sequelae of SARS-CoV-2”, “PASC”, “post-COVID-19 condition”, “PCC”, “post-acute COVID-19 syndrome”, “PACS”, “post-COVID syndrome”}). Template 2 anchored on mechanism vocabulary ({“viral persistence”, “antigen persistence”, “dysbiosis”, “microbiome”, “mast cell”, “small fiber neuropathy”, “autoantibody”, “microclot”, “mitochondrial”, “neuroinflammation”, “hypothalamic”}). Template 3 anchored on outcome vocabulary ({“fatigue”, “brain fog”, “post-exertional malaise”, “dysautonomia”, “POTS”, “dyspnea”, “cognitive impairment”}). Template 4 cross-joined Template 1 × Template 2 with operator AND. Deduplication used PMID as primary key with DOI fallback; records with neither stable identifier were resolved via title + first-author + year + journal exact-match against OpenAlex. The deduplicated yield (5,101) is an aggregate of 8,237 raw hits after dedup-collapse (38 % overlap rate across databases, comparable to other multi-database biomedical sweeps).

The sweep does not constitute a PRISMA-compliant systematic review at the screening layer (T6 PRISMA-2020 item 9 is the canonical gap; §10.2). Single-LLM screening is the screen-layer architecture; a human gold-standard dual-screener pass against a stratified sample is a v3.3 priority. The current architecture is sufficient to support the per-chain comparative-evidence claims in §4–§7, with the per-paper screening-layer caveat explicit. The honest-gap disclosure protocol (§3.10) propagates the limitation to the manuscript narrative throughout.

§3.2 Eligibility criteria

Records were eligible for atlas inclusion if they (a) reported empirical findings (cohort, case-control, RCT, MR, autopsy, biopsy, imaging, in-vivo animal model, ex-vivo tissue, or organoid) bearing on a Long COVID mechanism, (b) included either an LC cohort or a comparator cohort relevant to cross-domain validation (ME/CFS, POTS, PTLDS, post-Ebola, autoimmune syndromes, MIS-C, sepsis survivors), or (c) provided methodological framework (PRISMA, GRADE, ROBINS-I, Cinelli-Hazlett, Pearl do-calculus). Conference abstracts without subsequent peer-reviewed publication were excluded. Letters and editorials were excluded unless they reported novel data. Pre-prints were included with explicit oa_status: pre-print annotation but were not separately weighted in tier assignment.

The atlas is mechanistic in framing: a paper that reports an RCT in LC but does not connect its outcome to a postulated mechanism is included only if it is cited as evidence for or against a chain card. The 806-trial corpus reported in v2 (reports/manuscript-v2.md) is preserved as a parallel resource in the v3.2 atlas (analysis/network-meta/intervention-catalog.json) but is not the unit of synthesis in v3.2; the unit of synthesis is the mechanism link card.

§3.3 Per-card extraction protocol

Each of the 222 mechanism link cards encodes the following structured fields (see exhibit G21 for PRISMA-27 compliance assessment): link_id, root (chain identifier), from_node and to_node (the antecedent and consequent biological nodes), mechanism (the inferential bridge between antecedent and consequent in 1–3 sentences), quantitative_parameters (encoded effect sizes, CIs, p-values, sample sizes from primary citations; this is the c4 strict-rubric input), primary_citations (PMID list, typically 3–10 per card), evidence_tier, directionality (forward / reverse / bidirectional + per-direction evidence), cross_links (named upstream and downstream chain cards), limitations (the card author’s per-card honest gaps), falsification_trigger (concrete trial or experiment that would falsify the claim; this is the c7 strict-rubric input), sex_modulation, vaccine_status_modulation, variant_modulation (which SARS-CoV-2 lineage the card applies to), pediatric_applicability, and reinfection_modulation.

Cards mutated in Phase 7 also encode phase7_audit_status (DEMOTE / PROMOTE / HOLD), phase7_audit_score (the 0–8 strict-criteria total), phase7_audit_threshold_for_current_tier (≥6 for ESTABLISHED, ≥5 for CONSISTENT, ≥3 for EMERGING, <3 for SPECULATION), and evidence_tier_pre_phase7 (the v3.2 tier preserved verbatim — this is the v3.2-immutability hook).

Per-card extraction was performed by AI sub-agents under pre-registered schema discipline (docs/v3.2-PRD.md). The PRISMA 2020 item 10 (data items extracted) and item 22 (limitations) pass universally across all 13 chains in the T6 PRISMA audit (analysis/prisma/v3.2-prisma-summary.md). The atlas’s schema discipline is one of its methodological strengths; the underlying low PRISMA score reflects structural choices about search and curation (single-curator, post-hoc PMID inclusion, no PROSPERO registration), not extraction-layer failure (see §11.1).

§3.4 Risk-of-bias assessment

Risk of bias was assessed under the Cochrane ROBINS-I framework for non-randomized studies and the Cochrane RoB 2 framework for RCTs. The composite ROBINS-I score is encoded per primary citation in analysis/risk-of-bias/robins-i-scores-{AD,EH,IL,MP,QZ}.json (filename suffixes encode alphabetical partitioning of 451 papers scored across the 5,101 record corpus). The five RoB tiers are low, moderate, serious, critical, and no information. Of the 451 scored papers, 174 score serious and 39 score critical; these counts are reported in the per-chain RoB profile exhibit G22. The criterion c3 strict-rubric input is the composite ROBINS-I tier across all primary citations of a given link card — a card “passes c3” if no primary citation scores critical-RoB.

The ROBINS-I scoring is LLM-panelist-derived (three-panelist majority vote on title, abstract, journal tier, and year, with the panelist trace logged per paper). It is not a gold-standard human RoB pass. This is an honest gap explicitly disclosed (§10.1) and the v3.3 priority is a Wave-3 RoB pass with human gold-standard panelists on the H11 (110 unscored PMIDs) and H13 (98 unscored PMIDs) subsets. The chain-level RoB distribution (exhibit G22) is sufficient to support the qualitative comparative claims made in §5 and Supplementary Appendix B; it is not sufficient to support claims about absolute per-paper RoB at journal-publication quality.

ROBINS-I five-partition logic and panelist procedure. The 451 scored papers were distributed across five JSON shards alphabetically by first-author family name (AD = A–D, EH = E–H, IL = I–L, MP = M–P, QZ = Q–Z). Per shard, each paper was scored by three independent LLM panelists prompted with the seven ROBINS-I confounding-domain prompts (D1 confounding, D2 selection, D3 classification, D4 deviations, D5 missing data, D6 measurement, D7 reporting). Each panelist returned a domain-level rating (low, moderate, serious, critical, no information) plus a per-domain justification string; the composite ROBINS-I score is the worst per-domain rating across the seven domains (Cochrane convention). Inter-panelist agreement was scored as Cohen’s κ on the composite tier (mean κ = 0.71, range 0.62–0.84 across shards), which falls in the “substantial agreement” band. Disagreements were resolved by majority vote across panelists; the residual minority position is preserved in the per-paper JSON for audit transparency.

Why LLM panelists rather than human reviewers. A human dual-RoB pass at the 451-paper scale would require ≈ 1,800 hours of trained-reviewer time at the conservative 4-hour-per-paper convention. The Brainworks v3.2 budget did not support that and the project elected to instrument LLM-panelist RoB at v3.2 scale with explicit limitation disclosure rather than truncate the corpus to a human-doable size. The limitation is real (panelists may systematically miss specific risk-of-bias signals that require domain expertise, particularly in D1 confounding) and is the v3.3 Wave-3 priority. The methodological position is that an LLM-panelist RoB pass on 451 papers is more informative than no RoB pass at all, with honest disclosure of the panelist architecture.

§3.5 Evidence-tier assignment rubric (v3.2 + v3.2.1)

The v3.2 evidence-tier vocabulary is the canonical five-tier scale ESTABLISHED, CONSISTENT, EMERGING, SPECULATION, NULL (atlas pre-reg §7.1). Tier assignment in v3.2 was a mechanism-and-coherence-weighted heuristic: a card earned ESTABLISHED status when (a) it had a mechanistically continuous story from antecedent to consequent that matched textbook physiology, (b) two or more primary papers reported convergent observations, and (c) the card cohered with the 13-chain causal partition and the 27-mediator polytope architecture. This regime captured three real epistemic virtues — mechanism plausibility, convergent literature counts, and conceptual coherence with the model — that the Phase 7 strict rubric (see below) does not credit. We expand on the trade-off in Supplementary Appendix B.

The Phase 7 strict re-tier rubric (analysis/phase7-retier/phase7-retier-recommendations.md) is an eight-criterion peer-review-defensible filter. Each card is scored 0–8 against:

• c1 PI-independence — ≥ 2 distinct research groups demonstrated across primary citations.
• c2 prospective design — ≥ 1 prospective cohort or RCT among primary citations.
• c3 ROBINS-I composite ≤ moderate — no critical-RoB anchors among primary citations.
• c4 effect-size confidence interval — quantitative_parameters encodes a 95 % CI excluding null.
• c5 cross-domain coherence — T8 verdict ≥ PARTIAL/VALIDATED in ≥ 2 comparator literatures.
• c6 negative control — T10 chain NC verdict PASSES or PARTIAL on the within-LC tracer dimension.
• c7 falsifier named — falsification_trigger names a concrete trial or experiment (ideally with an NCT identifier).
• c8 sign consistency — directionality.forward_supported is consistent across cohorts.

The tier thresholds applied to the c1–c8 sum are: ESTABLISHED ≥ 6/8; CONSISTENT ≥ 5/8; EMERGING ≥ 3/8; SPECULATION < 3/8. Cards with intermediate UNVERIFIABLE scores (e.g., 5/8 with two UNVERIFIABLE) are scored conservatively — UNVERIFIABLE does not credit a True. The audit is therefore hostile-reviewer-safe: any DEMOTE verdict is defensible on the canonical data alone.

Score-to-tier mapping mechanics. The thresholds are deliberately staged so that ESTABLISHED is a high-bar tier (6/8 means at most two failed criteria, ideally not including c3 ROBINS-I or c7 named falsifier; cards lacking c3 or c7 cannot pass through to ESTABLISHED even at score 6/8 — see Supplementary Methods S2). Promotion is permitted (a card scoring 6/8 with v3.2 tier CONSISTENT promotes to ESTABLISHED); demotion is also permitted (a card scoring 4/8 with v3.2 tier ESTABLISHED demotes to CONSISTENT). The 80 HOLD cards are those whose strict-rubric score is within their v3.2 tier band without crossing any threshold. The 107 demotions distribute across the chain set in proportion to chain-level evidence-form deficits (Supplementary Appendix B (relocated §7.5), exhibit G36). The 35 promotions concentrate in chains with strong individual primary citations — H6 (mitochondrial), H10 (neuroinflammation), and H5 (autoantibody) — where the v3.2 conceptual-coherence tier was conservative.

The five tiers’ substantive meanings. ESTABLISHED is the publication-grade tier: a reviewer at Cell, Lancet, or Nature Medicine can accept the card on its own evidence without supplementary documentation. CONSISTENT is the survival-grade tier: the card is defensible against ordinary peer review but a hostile reviewer can challenge it on specific gaps. EMERGING is the working-hypothesis tier: the card is mechanistically plausible and has supporting primary citations but lacks the rigor for ESTABLISHED. SPECULATION is the candidate-hypothesis tier: the card encodes a mechanism that the field is actively debating but the evidence is preliminary or contested. NULL is reserved for explicitly-refuted claims (no v3.2.1 card lands in NULL; the refuted-claim register sits separately in the project documentation). The v3.2.1 distribution — 30 ESTABLISHED / 68 CONSISTENT / 98 EMERGING / 26 SPECULATION — maps the corpus’s evidence-form profile onto these substantive tiers.

The Phase 7 audit is applied to all 222 cards. The resulting tier distribution is reported in Supplementary Appendix B (relocated §7.2) (exhibit G36); the chain-by-chain reinterpretation in Supplementary Appendix B (relocated §7.3); the 30 ESTABLISHED survivors in Supplementary Appendix B (relocated §7.5) (exhibit G38); the four chronic gates in Supplementary Appendix B (relocated §7.4) and §8 (exhibit G37).

§3.6 Network construction — chains, mediators, polytope, inter-root pairs

The v3.2 atlas is structured as a partially-cyclic directed graph over 13 chains, 27 convergent mediator nodes, and a 117-cell symptom polytope (13 chains × 9 symptom domains). The 9-symptom polytope is clinical-observation-anchored (RECOVER PASC scale, MY-LC discriminative panel, BIONIC neurocognitive battery). The chain architecture (R1, R2, R3 + ten H-chains) is the deliverable of the Phase 6 T11 capstone synthesis (analysis/synthesis/v3.2-phase6-synthesis.md §5.4, CF-1). The 27 mediator nodes (X.Dysbiosis, X.SCFA, X.Histamine, X.Tcell_exhaustion, X.Complement, X.IL6, X.TNFa, X.IL1b, X.Microclot, X.Spike, X.EBV_lytic, X.Autoantibody, X.NETs, X.Cortisol, X.CRH, X.SympNS, X.ParaNS, X.Endothelium, X.BBB, X.Microglia, X.Astrocyte, X.MitoCDR, X.NAD, X.MAVS, X.MastCell, X.Vagus, X.IFNI) emerge from cross-chain convergence analysis on the link cards’ to_node and from_node fields. Each mediator has typed direct upstream and downstream chain anchors (see exhibit G16).

The 156-ordered-pair inter-root verdict matrix (analysis/inter-root-causal-graph/directed-pairs-{Y1,Y2}.json; exhibit G18) is the deliverable of Phase 5b T12. For each ordered (chain_i → chain_j) pair across the 13 × 12 = 156 ordered pairs (the diagonal is omitted as a chain is not its own cause), a verdict is assigned from PROBABLE_CAUSE, PLAUSIBLE, INSUFFICIENT_EVIDENCE, EQUIPRIMORDIAL, or RULED_OUT based on direct PMID-anchored mechanism evidence in the source and target chain cards plus cross-link annotations. The combined Y1 + Y2 verdict distribution is 33 PROBABLE_CAUSE / 58 PLAUSIBLE / 56 INSUFFICIENT_EVIDENCE / 9 EQUIPRIMORDIAL / 0 RULED_OUT. Verdicts are reported per pair with PMID anchors in the JSON; the chain-level out-degree distribution is plotted in G18.

Cycle annotation (exhibit G19, analysis/inter-root-causal-graph/cycles.md, cycles.json) admits the one true PROBABLE_CAUSE bidirectional cycle (H7 ↔︎ H12) and four asymmetric near-cycles (forward PROBABLE_CAUSE, reverse PLAUSIBLE): H10 → H7, H8 → H10, H10 → H12, H7 → H8. Two additional asymmetric pairs (H5 → H10, H6 → H10) surface in the directed-pair data scan; the synthesis treats H5 → H10 as the fifth near-cycle on the chronic-tail surface and H6 → H10 as a hierarchical (not cyclic) edge. The cycle layer is the empirical signature that the atlas is not a strict DAG at the H7/H8/H10/H12 central-amplifier cluster; the T1 Bayesian network (§3.7) drops these cycles by construction (the noisy-OR inference machinery is DAG-only), and the cycle annotation is the manuscript’s transparent representation of what the T1 DAG cannot natively encode.

Per-pair verdict procedure. For each of the 156 ordered (chain_i → chain_j) pairs, three LLM analyst-panelists were prompted with the source-chain’s mechanism summary, the target-chain’s mechanism summary, the candidate primary-citation evidence linking them, and a five-tier verdict rubric: PROBABLE_CAUSE (direct PMID-anchored mechanism evidence ≥ 1 prospective or RCT design plus mechanism plausibility ≥ 0.85), PLAUSIBLE (mechanism plausibility ≥ 0.70 with at least one cross-link in atlas evidence), INSUFFICIENT_EVIDENCE (mechanism plausibility 0.30–0.70, no direct supporting citation), EQUIPRIMORDIAL (the pair forms a bidirectional cycle with comparable evidence in both directions), and RULED_OUT (active disconfirming evidence). Each panelist returned a verdict plus a rationale string; the per-pair canonical verdict is the majority across panelists with ties broken toward the more conservative verdict (PLAUSIBLE > PROBABLE_CAUSE; INSUFFICIENT > PLAUSIBLE). The audit trail is preserved per pair in directed-pairs-{Y1,Y2}.json.

§3.6.1 Pearl identifiability framework and why most NMA effect estimates fail Pearl

The atlas’s T9 counterfactual layer rests on Judea Pearl’s do-calculus identifiability framework. The framework treats the question “what is the causal effect of intervention X on outcome Y?” as separable into a statistical question (what is P(Y | X) in the observed data?) and a causal question (what is P(Y | do(X)) in the interventional distribution?). The interventional distribution is identifiable from observational data if and only if a sufficient adjustment set can be constructed under either the back-door criterion (a set of pre-intervention covariates that blocks all back-door paths from X to Y) or the front-door criterion (a measured mediator that satisfies the front-door admissibility conditions). When no identifying adjustment exists, the causal effect is unidentifiable — not merely uncertain, but provably outside the inferential reach of the observed data.

Why this matters for LC therapeutic claims. Most LC mechanism papers report observational effect estimates that condition on incomplete adjustment sets. The unmeasured confounders enumerated in the T9 latent-confounder list (U_HLA, U_diet_lifestyle, U_sex_hormones, U_HSV_susceptibility, U_individual_vagal_tone, U_baseline_microvascular_health) leave open back-door paths in the majority of the catalogued interventions. Of ten T9-analysed candidates, four are unidentifiable (anti-C5, T-cell-exhaustion modulation, anti-spike passive transfer, microclot apheresis at the do-X operator level), three are identifiable-but-thin (IVIG-SFN, butyrate supplementation, fludrocortisone), and three are identifiable-and-supported (SIM01, H1+H2 blockade, the contested H3.L10 in PR-1 disposition).

The RCT design is the textbook back-door identification: random assignment severs the back-door paths by construction. This is the structural reason the SIM01 RCT (Lau 2024 PMID 38071990) satisfies Pearl’s back-door criterion directly while observational X.Dysbiosis-to-symptom-relief associations do not. Pearl identifiability is the methodological bar the atlas places between mechanism strength (which many chains have) and causally-defensible therapeutic recommendation (which few interventions earn). The bar is high by design and is what distinguishes the T9 counterfactual layer from the T7 NMA: the NMA pools observational and RCT effect estimates under a meta-analytic framework that does not require Pearl identifiability, and the resulting effect estimates are correctly characterized as associational under the meta-analytic model rather than as causally identified do-X effects.

The atlas’s epistemic posture is to report both: the T7 NMA estimates as the field’s pooled meta-analytic position, and the T9 do-calculus identifiability verdicts as the Pearl-defensible interventional position. The two should not be conflated. The headline therapeutic finding (§7) is anchored to T9, because SIM01 is the one intervention where the conditions align.

§3.7 The eleven analytical techniques

The v3.2 evidence layer is the union of eleven independently executed analytical techniques. They are summarized here; their per-technique results are reported in §4–§6.

• T1 Bayesian noisy-OR network (PyMC NUTS MCMC + Beta-Binomial conjugate analytic; analysis/bayesian-net/). A 482-edge canonical DAG with chain-typed inter-root edges and mediator-knockout sensitivity analysis. R-hat = 1.0000, ESS_min = 6,442, 0 divergences; posterior predictive r = 0.984 (with honest caveat about Beta-Binomial conjugacy yielding automatic PPC accuracy — see §11.3). Composite mediator-influence ranking via in-shift + out-swing metric (X.Dysbiosis = 0.234, X.SCFA = 0.114, X.Histamine = 0.061, X.Tcell_exhaustion = 0.049, X.Complement = 0.048). Posterior topology rendered as exhibit G23; PPC as G24.
• T2 Mendelian randomization (analysis/mr/v3.2/). Four instruments (ABO B-allele rs8176719; IL6R rs2228145 + 2 LD-r²>0.95 proxies → effectively single Wald-ratio; HLA-region; ACE2 eQTL); five outcomes (LC, fatigue, brain fog, dysautonomia, dyspnea), Bonferroni α = 7.7 × 10⁻⁴. F-statistic minimum = 27.6 (passes weak-instrument threshold). Outcome GWAS: Lammi 2025 Nature Genetics PMID 40399555 (6,450 LC cases, 1.09M controls). Bonferroni-robust positives: IL6R → fatigue β ≈ +1.5 (single-instrument Wald p ≈ 0.001–0.003). MR forest plot rendered as exhibit G27.
• T3 Cinelli-Hazlett robustness (analysis/sensitivity/cinelli-hazlett-bounds.json). Closed-form robustness value RV = 0.5 × (√(f⁴ + 4f²) − f²) where f² = t²/df, computable on 9 of 58 ESTABLISHED-tier cards that report regression t-statistic + df; structured heuristic RV for the remaining 49. Distribution: 14 HIGH_ROBUST / 29 MODERATE / 13 LOW post-Phase-6 reconciliation. Rendered as exhibit G25.
• T4 per-link polytope leverage (analysis/sensitivity/per-link-polytope-loss.json). 222 cards × 117 polytope cells; per-card single-edge cut analysis. Four post-Phase-6 SPOFs (single points of failure): H3.L10, H10.L18, H11.L13, H11.L16 (= 4/222 = 1.8 %). Rendered as exhibit G26.
• T5 adversarial chain-vulnerability scoring (analysis/redteam/chain-vulnerability-scores.json). Simulated three-panelist hostile peer review per chain; vulnerability score 0–10. H6 = 8.5 (most robust), H10 = 8.0, R1 = 7.5; H11 = 3.0 (weakest, by intentional design — the chain that red-teams the field’s single-driver hypotheses is itself rated weakest); H12 = 5.5, H5 = 5.5. Rendered as exhibit G28.
• T6 PRISMA 2020 audit (analysis/prisma/). All 13 chains low-tier (mean 11.0/27); 8 items score no universally. Exhibit G21.
• T7 network meta-analysis topology (analysis/network-meta/). 42 catalogued interventions; 9 placebo-bridged edges; 0 head-to-head intervention-vs-intervention trials; 0 closed loops → no consistency test → no defensible SUCRA/P-score. Star topology rendered as exhibit G33; per-class effect estimates G34.
• T8 cross-domain coherence (analysis/cross-domain/per-chain-per-domain-matrix.json). 13 chains × 5 comparator literatures = 65 cells. Distribution: 19 VALIDATED / 19 PARTIAL / 21 INSUFFICIENT / 6 LC_SPECIFIC. Strongest comparator: ME/CFS at 8/13 chains VALIDATED-or-PARTIAL. Exhibit G29.
• T9 counterfactual simulation (Pearl do-calculus). Ten highest-leverage causal targets (5 from T4 SPOFs + 5 from T1 mediator hubs); back-door + front-door identifiability; latent-confounder enumeration (6 confounders: U_HLA, U_diet_lifestyle, U_sex_hormones, U_HSV_susceptibility, U_individual_vagal_tone, U_baseline_microvascular_health). Intervention-priority calculation P = E × I × V (effect-weighted, identifiability, validation). SIM01 P = 9.00 (rank 1); H3.L10 IMC-2+Paxlovid (Pridgen combination antiviral) P ≈ 3.75 (rank 2, Phase 11d re-score + Phase 11e cohort-disclosure addendum; CONTESTED-POSITIVE pending HEAL-LC NCT07597902 readout); X.Histamine P = 1.96 (rank 3); IVIG-SFN P = 1.82 (rank 4); X.SCFA P = 1.12 (rank 5); anti-C5 P = 0; X.Tcell_exhaustion P = 0. Exhibits G31 (counterfactual forest) and G32 (per-symptom relief).
• T10 negative-control battery (analysis/negative-controls/per-chain-negative-controls.json). Lipsitch/Tchetgen-Tchetgen NCE / NCO / tracer applied per chain. 7/13 chains partial-fail the cross-syndrome 12-month NCE; only H6 passes outright (Appelman/Wüst 2024 PMID 38381103 activity-matched CPET-PEM design). Exhibit G30.
• T12 inter-root directed pairs (described above, §3.6). 156-pair verdict matrix, two-wave (Y1 + Y2). Exhibit G18.
• T11 capstone synthesis (analysis/synthesis/v3.2-phase6-synthesis.md). Triangulates T1–T10 + T12 into convergent findings (CF-1 through CF-11) and arbitrates technique-vs-technique disagreements (D-1 through D-9; disagreement-table.md). The disagreement arbitration is reported in §5.9 of this manuscript.

§3.8 Phase 7 strict-criteria re-tier audit

The Phase 7 audit (analysis/phase7-retier/phase7-retier-master-audit.json, commit c0d5934) applied the eight-criterion strict rubric (§3.5) to every link card. The audit was executed in a single pass over the 222 cards; each card receives a 0–8 score, a recommendation (DEMOTE / PROMOTE / HOLD), and a per-criterion verdict trace. The audit is conservative (UNVERIFIABLE does not credit a True), reproducible from the canonical JSON inputs, and machine-auditable.

The audit produced 107 demotions, 35 promotions, and 80 holds. Of the 35 promotions, 26 land in ESTABLISHED; of the 107 demotions, 22 land in SPECULATION. The pre/post fleet distribution is reported in Supplementary Appendix B (relocated §7.2) (exhibit G36) and the four chronic gates (the criteria that fail to credit at 59–96 % across the demotion set) are reported in Supplementary Appendix B (relocated §7.4) and developed as a field diagnosis in §8 (exhibit G37). The 30 ESTABLISHED survivors are listed by chain in Supplementary Appendix B (relocated §7.5) (exhibit G38).

§3.9 Dual-version preservation

The v3.2.1 strict-criteria-compliant canon was produced by Phase 7.5 by cloning the v3.2 atlas directory tree and applying the 142 Phase 7 mutations to the clone while leaving the original v3.2 cards byte-untouched. The v3.2-pathways/ parent directory is locked by an _immutable_canonical.json marker; the v3.2.1-pathways/ child carries an _provenance.json that names the parent path, the audit commit, and the rubric. Every mutated v3.2.1 card preserves the v3.2 tier verbatim in an evidence_tier_pre_phase7 field. Every HOLD card in v3.2.1 is byte-identical to its v3.2 counterpart except for an additive phase7_audit_status = 'HOLD' annotation block.

The two-canon architecture is the Captain-directed transparent representation of the gap between conceptual-coherence-weighted curation (v3.2) and per-card strict peer-review defensibility (v3.2.1). Neither rubric is wholly right; both have real epistemic content. The manuscript cites both per claim with the discipline noted in Supplementary Appendix A (relocated §1.4). Downstream consumers — Phase 8 exhibits, this manuscript, and the Phase 10 HTML render — must declare per-figure or per-paragraph which canon underwrote the claim.

§3.10 Honest-gap disclosure protocol

The atlas is constructed under an honest-gap-preferred policy: when an analytical technique returns INSUFFICIENT, UNVERIFIABLE, or a contested verdict, the gap is preserved in the canonical record rather than synthesized through. Honest gaps are propagated to §11 of this manuscript and to per-card limitations fields. The disagreement table (disagreement-table.md) catalogues nine technique-vs-technique disagreements (§5.9), six of which dissolve on closer inspection as “different questions” rather than “different answers”; the three substantive disagreements (D-1 T1 vs T12 on cycles, D-5 T1 vs T9 on H3.L10, D-8 H12 reclassification threshold) are reported transparently with their arbitration verdicts.

The atlas’s honest-gap discipline is one of its methodological strengths and is the single most important feature for peer-review defensibility: a reviewer who flags an unaddressed gap can be answered with the per-card limitation field, the synthesis honest-gap entry, or the §11 limitations section, all of which are pre-existing canonical artifacts rather than rebuttal-stage inventions.

---

§4 Architecture results

The master architectural integration of all 13 chains, with edge thickness encoding T1 Bayesian-network posterior directional-support weights, is shown in exhibit G57 — the manuscript’s capstone figure. G57 layers the three roots (sized by outgoing-edge count), the ten amplification hubs (bordered by tier status), the twenty-seven mediators (color-graded by composite-influence rank), and the nine symptom domains into a single integrated DAG. The H7⇌H12 self-sustaining cycle is highlighted in red; the SIM01 keystone (R2 → X.Dysbiosis, Pearl-defensible P = 9.00) is annotated in gold. Exhibit G57 (capstone, repeated for reference): Integrated 13-chain causal-chain architecture with relative directional-support weights. Edge thickness = T1 Bayesian posterior PROBABLE_CAUSE weight; node size = MCMC composite influence ranking; edge color = chain-of-origin (NAVY R1, ACCENT_BLUE R2, ELECTRIC_BLUE R3). The H7⇌H12 cycle and SIM01 keystone are annotated.

Sub-section results follow.

§4.1 Three roots, ten amplification hubs (CF-1)

The v3.2 atlas resolves into a three-root + ten-amplification-hub architecture (exhibit G15; integrated view in G57; per-chain detail in G40-G52). The three causal roots are:

• R1 Viral persistence — sustained SARS-CoV-2 antigen, replicative reservoir, sgRNA detection, and circulating S1 spike protein. Anchored by Stein 2022 cross-tissue autopsy PMID 36517603, Swank / Walt circulating S1 longitudinal proteomics, Goh 2022 gut-biopsy reservoir detection PMID 34248921, Greene 2024 BBB-MRI disruption PMID 38388736.
• R2 Microbiome dysbiosis — altered gut microbial composition, reduced short-chain fatty acid (SCFA) producer abundance, increased Gram-negative LPS translocation, vagal afferent sensitization. Anchored by Liu 2022 Gut PMID 35082169, Zhang 2022 Gastroenterology SCFA biosynthesis impairment PMID 34687739, and the Lau 2024 SIM01 RCT PMID 38071990.
• R3 Endothelial / coagulopathy — endothelial dysfunction, von Willebrand-factor / ADAMTS13 imbalance, complement-coagulation cross-talk, microclot biology. Anchored by Cervia-Hasler 2024–2025 complement-coagulation findings PMIDs 39723824 and 39942772, and the Pretorius / Kell microclot literature.

These three chains have strong PROBABLE_CAUSE outgoing reach into the H-chains in the 156-pair Y1 + Y2 verdict matrix: R1 → H2, H5, H6, H7, H8, H10, H13 are all PROBABLE_CAUSE; R2 → R3, R2 → H10 are PROBABLE_CAUSE; R3 → H6, H7, H10, H13 are PROBABLE_CAUSE. The H-chains do not reach back into R-chains with PROBABLE_CAUSE evidence. The structural conclusion is that R1, R2, R3 are the upstream causal drivers, and the ten H-chains (H2 EBV-autoimmunity, H3 composite antigen amplification, H5 autoantibody axis, H6 mitochondrial CDR, H7 autonomic / small-fiber neuropathy, H8 mast-cell hyperactivation, H10 central neuroinflammation, H11 musculoskeletal / Groysman framework wrap, H12 HPA axis, H13 pediatric) are downstream amplification arcs.

This architecture supersedes the v3.1 twelve-root framing (CF-1; expanded-analysis-v3.2-vs-v3.2.1.md). The empirical signal that demands the move is the asymmetric out-degree pattern: R1, R2, R3 are upstream-dominant; the H-chains are downstream-dominant. T1 (Bayesian network root-to-mediator posteriors), T7 (NMA per-chain RCT distribution — concentrated in R1 acute prevention), and T5 (chain vulnerability scoring — R1 score 7.5, R2 and R3 also among the more robust) corroborate. The architecture is [ESTABLISHED] per the atlas pre-reg §7.1 vocabulary and survives the Phase 7 strict-criteria audit at the architectural-claim level (no per-card mutation can change the 13-chain partition; the partition is a curation-layer decision, not an evidence-layer claim).

The architecture is rendered in exhibit G15 with R1, R2, R3 in the upstream tier, the ten H-chains in two mid-stream tiers (H2/H3 directly downstream of R1; H5, H6, H7, H8, H10, H12 in the central amplifier cluster; H11 and H13 as meta and pediatric specializations), and the 117-cell symptom polytope on the right.

§4.1.2 Per-chain narrative panel

The atlas’s thirteen chains each have a distinct biological identity, evidence profile, and atlas role. They are summarized here at ~100 words per chain so the architecture can be read at a glance without re-tracing the prior literature.

R1 — Viral persistence. Sustained SARS-CoV-2 antigen, replicative reservoir, sgRNA detection, and circulating S1 spike protein at 6–18+ months post-acute infection. Anchored by Stein 2022 cross-tissue autopsy PMID 36517603, Goh 2022 gut biopsy persistence PMID 36131932, Proal & VanElzakker 2021 PMID 34248921 review, Proal et al 2023 Nat Immunol PMID 37667052, and Swank longitudinal S1 proteomics PMID 39389851. R1 is the most-supported chain in the prevention frame (Paxlovid acute HR ≈ 0.74, three replications) and the principal upstream driver in the hierarchical model. R1 is also the chain that most clearly exhibits cross-syndrome distinctiveness — the sustained S1 antigenemia signature is the only mechanism unparalleled in the comparator literatures.

R2 — Microbiome dysbiosis. Altered gut microbial composition (reduced Bifidobacterium, Faecalibacterium, Roseburia; increased Gram-negative pathobionts), reduced SCFA-producer abundance, increased LPS translocation, and downstream vagal afferent sensitization. Anchored by Liu 2022 Gut PMID 35082169, Zhang 2022 Gastroenterology PMID 34687739, and the Lau 2024 SIM01 RCT PMID 38071990. R2 is the chain carrying the keystone Pearl-defensible therapeutic claim of the atlas. The R2 backbone (LPS-translocation → systemic-inflammation → vagal afferent activation) is among the most cross-domain-validated cards in the atlas (ME/CFS gut-dysbiosis literature confirms the pathway in an independent post-viral syndrome).

R3 — Endothelial dysfunction and coagulopathy. Endothelial activation (vWF / ADAMTS13 imbalance), complement-coagulation cross-talk, microclot biology, and microvascular ischemic burden. Anchored by Pretorius / Kell microclot literature (PMID 33203441 prevalence; recent extensions PMIDs 39723824 fibrinogen-self-assembly, 39942772 clot proteome, 39434957 fibrinaloid activity), Cervia-Hasler complement-coagulation findings, and the Charfáuddin / Pretorius HELP-apheresis observational cohorts. R3 is intervention-deplete at the LC-specific RCT level (no completed sham-controlled HELP-apheresis trial; FORTRESS and recommended sham-RCT designs are catalogued in §12). R3’s chain-vulnerability rank is mid-to-high; the microclot-as-mechanism debate is unresolved at the bedside but mechanistically supported.

H2 — EBV reactivation × T-cell-exhaustion axis. EBV lytic reactivation in post-acute COVID, with T-cell exhaustion (PD-1, TIGIT, LAG-3 elevation) on EBV-specific and bystander populations. Anchored by Gold 2021, Su 2022 (Heidelberg multi-omics PMID 35216672), and Klein 2023 PMID 37748514. H2 is the chain that connects R1 sustained antigen to broader herpesvirus reactivation; the mechanism is supported in the multi-omics MY-LC cohort. H2 carries 2 ESTABLISHED-tier cards after Phase-7 audit and is one of the chains most likely to gain ESTABLISHED-tier promotions in v3.3 as additional cohort replications land.

H3 — Composite-antigen amplification. A meta-chain that integrates spike, S1 fragments, and host autoantigens into a composite antigen-amplification mechanism via IFN-I and complement. Anchored by Phetsouphanh 2022 IFN-I signature work and the Yang / Stein-Heider-Schmidt complement-coagulation cross-talk literature. H3 carries the H3.L10 SPOF card in PR-1 disposition: the underlying mechanism is plausible but the strict-rubric c4 (CI excluding null) gate fails because the supporting cohort’s effect estimates report on subset-stratified outcomes rather than primary-endpoint composite contrasts. H3 is the chain that the contested IMC-1 and HEAL-LC trials will most directly inform. The H3 chain is broader than the FORTRESS IMC-1 monotherapy card: the clinically-deployed Pridgen protocol is IMC-2 (valacyclovir + celecoxib) ± Paxlovid — a combination-antiviral regimen that is empirically and mechanistically distinct from IMC-1 monotherapy (dual-antiviral via the COX-2-dependent herpesvirus-replication pathway plus SARS-CoV-2 Mpro inhibition), and the FORTRESS null is best read as informative about the monotherapy card rather than as a falsification of the combination-protocol mechanism story (§6.5).

H5 — Autoantibody axis. GPCR-targeting autoantibodies (β1-AR, M2-AChR, AT1R, ETB receptors) and broader autoantibody-driven autoimmunity. Anchored by Wallukat 2021 PMID 33880442, the BHC-202 trial program targeting GPCR autoantibodies, and the Vernino IVIG-SFN trial PMID 38311655 in the autoantibody-positive SFN subset. H5 carries 5 ESTABLISHED-tier cards under Phase-7 audit — the strongest autoantibody-axis evidence in the atlas. The chain’s clinical implication is the GPCR-antibody-stratified subset hypothesis: BHC-202 reads out as the most informative pending trial.

H6 — Mitochondrial cell-danger response. Persistent mitochondrial dysfunction, cell-danger-response activation, NAD⁺ depletion, and the activity-matched CPET-PEM signature. Anchored by Naviaux 2020 PMID 31877376 (CDR framework), Wallukat 2020 PMID 32977857 (autoantibody-mitochondrial cross-talk), Appelman 2024 Nat Commun PMID 38177128 (the gold-standard activity-matched CPET-PEM design), and now Tasoula 2026 Front Immunol doi:10.3389/fimmu.2026.1776555 — the field’s first integrated multi-omics (transcriptomic + proteomic + metabolomic) cross-species (Syrian hamster + human) cross-tissue (skeletal muscle, cardiac, kidney, lung, multi-region brain, PBMCs, serum) longitudinal (acute → 12+ months post-infection) characterization of the OXPHOS-suppression signature, which extends the mitochondrial-CDR mechanistic frame from the cohort level into peer-reviewed molecular convergence and directly compares PCS-CFS vs T2bFA-CFS skeletal muscle biopsies at the transcriptional level — the strongest published molecular evidence of PCS↔︎ME/CFS mitochondrial convergence to date (see §9.3.3). H6 has the lowest mean ROBINS-I severity (1.25) and is the only chain to pass the cross-syndrome NCE outright. H6 carries 7 ESTABLISHED-tier cards and is the most-robust chain in the atlas. The chain’s mechanistic specificity is further sharpened by Tasoula 2026’s proposed cascade — acute OXPHOS suppression → HIF-1α stabilization / Warburg-like glycolytic shift → mROS elevation → mtDAMP release (mtDNA, mtdsRNA, cardiolipin) → pattern-recognition-receptor (PRR) activation → chronic interferon / cytokine signaling — which provides the molecular substrate for the H6-→-H10 hierarchical edge (§4.5) and for the cross-chain mtDAMP bridge to H2 and R1 (§11).

H7 — Autonomic / small-fiber neuropathy. Small-fiber neuropathy (reduced epidermal nerve-fiber density), POTS, dysautonomia, and brainstem-mediated cardiovagal dysregulation. Anchored by Oaklander 2022, the H7-H8 mast-cell cross-talk literature (Theoharides 2017 PMID 28987303), and the H7 ↔︎ H12 cycle integration. H7 carries 3 ESTABLISHED-tier cards. The chain forms the central H7 ↔︎ H12 cycle and is the principal mediator of the dysautonomia, POTS, and orthostatic-intolerance phenotype.

H8 — Mast-cell hyperactivation. MCAS-like phenotype in LC, with elevated tryptase and histamine, multi-domain symptoms (GI, sensory, dermatologic). Anchored by Afrin 2020 PMID 32920235, Theoharides 2017 PMID 28987303, and the H8 ↔︎ H7 mast-cell-autonomic cross-talk. H8 carries 4 ESTABLISHED-tier cards and is the chain underlying the combined-mast-cell-stabilizer RCT proposal (§12). H8’s strict-rubric profile is good: the underlying tryptase / histamine cohort data report effect estimates with confidence intervals, and prospective cohort designs exist.

H10 — Central neuroinflammation. Microglial activation (TSPO-PET), blood-brain barrier disruption, sustained circulating spike and S1, and central inflammatory cytokine signaling. Anchored by Greene 2024 Nat Neurosci PMID 38388736 (BBB disruption), Braga 2023 PMID 36571857 (TSPO-PET imaging), and Klein 2023 immune-profiling, with supporting cross-chain molecular substrate from Tasoula 2026 Front Immunol doi:10.3389/fimmu.2026.1776555 (region-specific cortical OXPHOS repression and persistent neuroinflammatory transcripts in hamster mPFC + human frontal-cortex autopsy samples; supplies an H6 → H10 mitochondrial-DAMP → microglial-activation mechanistic bridge). The neuroinflammation-as-unifying-mechanism synthesis is further contextualized by Martins 2025 Preprints.org doi:10.20944/preprints202506.1469.v1 — a comprehensive narrative synthesis identifying GFAP, sTREM2, S100β, IL-6, TNF-α (fluid) and TSPO-PET, ASL-MRI, MR spectroscopy (myo-inositol, choline), diffusion MRI (imaging) as candidate diagnostic biomarkers, with novel mechanistic candidates including local sleep intrusions, impaired memory reconsolidation, and astrocyte-mediated network destabilization. Martins 2025 is a preprint (not peer-reviewed); under the Phase 7 strict rubric it cannot contribute to chain-tier promotion (fails c2 prospective-design and c8 peer-review-equivalent gates) and is cited here only as a synthesis-context anchor per the §3.10 honest-gap disclosure protocol — the named biomarkers and novel pathways are pre-registered as v3.3 hypothesis-generating items (§12.4), not v3.2 evidence-promotion entries. H10 has 7 ESTABLISHED-tier cards — tied with H6 for the largest survivor count. H10 forms the central-amplifier cluster with H7, H8, H12 and is the principal mediator of the brain-fog and cognitive-dysfunction phenotype.

H11 — Flat-network framework wrap. The Groysman 2026 Frontiers in Hypothesis & Theory framework included for systematic red-teaming. H11 fails the C1–C8 rubric universally (§4.7), and all five framework cards (H11.L02–L05, H11.L20) demote to SPECULATION in Phase-7 audit. The H11 biological insight (LC is multi-mechanism with feedback loops) is empirically supported and subsumed into the hierarchical-DAG-with-annotated-cycles model. H11 carries 110 unscored PMIDs (highest in atlas) and 7 % ROBINS-I coverage; the chain is the v3.3 Wave-3 RoB-pass priority.

H12 — HPA-axis exhaustion. Hypocortisolism, blunted ACTH response, chronic CRH dysregulation. Anchored by Klein 2023 PMID 37748514 (MY-LC hypocortisol cohort) and cross-domain validated against autoimmune hypopituitarism (Sheehan’s; PMID 18230820 pending verification). H12 is reclassified from “parallel root” to “central downstream amplifier” (§4.6, CF-7). H12 forms the bidirectional cycle with H7 and is the principal HPA-axis hub feeding back onto autonomic regulation. H12 carries 2 ESTABLISHED-tier cards.

H13 — Pediatric LC. Pediatric-distinctive Long COVID, including MIS-C-related sequelae, post-acute pediatric immune dysregulation, and developmental-trajectory effects. Anchored by Buonsenso 2022 / 2023 series and the broader pediatric LC literature. H13 carries 0 ESTABLISHED-tier cards under Phase-7 audit (98 unscored PMIDs is the highest unscored count, and 0 % ROBINS-I coverage). The chain is the v3.3 Wave-3 RoB-pass priority alongside H11 and is the chain whose evidence-form gap is most clinically consequential (pediatric LC trial design is structurally distinct from adult).

§4.2 Twenty-seven mediator convergence (G16)

The 222 link cards converge onto 27 mediator nodes that recur as the to_node of upstream cards and the from_node of downstream cards. The mediator-convergence map is rendered as exhibit G16. The composite influence ranking in the T1 MCMC sensitivity analysis is dominated by five mediators:

Rank	Mediator	Composite influence	Primary upstream chains	Primary downstream symptoms
1	X.Dysbiosis	0.234	R1 (via gut reservoir), R2 (direct), H8 (mast-cell-mediated mucosal effect)	GI, fatigue, brain fog, sleep
2	X.SCFA	0.114	R2, X.Dysbiosis	Fatigue, cognitive, immune
3	X.Histamine	0.061	H8, X.MastCell, R2 (via LPS-MC activation)	Brain fog, flushing, GI, fatigue
4	X.Tcell_exhaustion	0.049	H2, R1 (sustained antigen), H5	Recurrent infection, fatigue
5	X.Complement	0.048	R3, H5 (IC-driven), H10 (CNS)	Microvascular, neuroinflammation

The X.Dysbiosis + X.SCFA “gut axis” combined influence (0.348) exceeds the cumulative composite influence of all other top-5 mediators combined. This is the empirical signature behind the SIM01 → X.Dysbiosis Pearl-defensible therapeutic claim (§6, §7). The analytic conjugate Beta-Binomial knockout-impact ranking is different (X.Tcell_exhaustion 0.0112, X.SympNS 0.0108, X.Endothelium 0.0064, X.MAVS 0.0031, X.IL1b 0.0026) because it asks a different question — “which mediators, if surgically removed, most damage the strongest-path probability” — rather than “which mediators concentrate the most causal information flow.” Both rankings are correct for their respective questions (T11 disagreement D-3); we use the MCMC composite for mediator-hub argument and the analytic knockout-impact for intervention-prioritization (§5.1, §6.2).

The mediator convergence layer demonstrates that the 13 chains are not biologically independent. They share substantial machinery: H2 and H5 both project onto X.Autoantibody and X.Tcell_exhaustion; H6, H10, and H12 all project onto X.MitoCDR and X.Cortisol; H7 and H8 share X.MastCell, X.Histamine, X.Vagus. This shared-machinery picture is the structural reason that the cross-syndrome non-specificity finding (§9; CF-3) holds: the mediator pathways are post-viral / chronic-illness final-common-pathways, and the LC-distinctive layer is concentrated at the upstream antigen-persistence node (R1 spike + S1 antigenemia, anchored by Stein 2022 PMID 36517603 and Swank S1 longitudinal proteomics; the only fully LC-specific cells in the cross-domain matrix, §4.8). The hierarchical-DAG-with-annotated-cycles architecture absorbs this picture by separating the upstream initiation layer (R1-dominated, COVID-specific) from the maintenance layer (mediator-cycle-rich, post-viral-shared).

§4.3 The 117-cell symptom polytope (G17)

The symptom polytope is the 13 chains × 9 symptom domains = 117-cell matrix that maps each chain’s evidence onto each clinical symptom domain. The nine domains are: fatigue (D1), post-exertional malaise / PEM (D2), brain fog / cognitive dysfunction (D3), dysautonomia / POTS (D4), gastrointestinal symptoms (D5), dyspnea / cardiopulmonary (D6), neuropathic pain / small-fiber symptoms (D7), sleep disturbance (D8), and mood / neuropsychiatric (D9). The polytope is rendered as exhibit G17.

Each polytope cell is populated by zero or more link cards whose to_node includes the symptom-domain identifier. The polytope is redundant by design: the mean number of reaching paths per cell is greater than five, and 34 of 117 cells (29 %) have ≥ 5 reaching paths. Single-edge cuts (T4 polytope leverage) identify the four post-Phase-6 SPOFs (H3.L10, H10.L18, H11.L13, H11.L16; §5.3) where the atlas’s polytope coverage in a specific cell currently depends on a single link card. Mean single-edge cut impact is 0.46 polytope cells per card (i.e., most cards’ removal leaves the polytope coverage map essentially unchanged).

The polytope coverage does not span the entire 117-cell space. Several cells (e.g., H11 / sleep, H13 / pain, R3 / mood) are partly or wholly empty in v3.2 — honest gaps preserved in the canonical record rather than synthesized through. The polytope coverage analysis is reported in analysis/sensitivity/per-link-polytope-loss.json and visualized in G17. The gaps are flagged as v3.3 priority and discussed in §11.4 and §12.1.

The polytope’s redundancy is the central structural reason that no single mediator is bottleneck-critical (CF-2, §5.1): the parallel-path structure absorbs the loss of any single edge. The four SPOFs (1.8 % of cards) are the exceptions; even there, the SPOF status is graph-theoretic (one polytope cell currently depends on the card) and not biological-bottleneck (the broader chain support is unaffected by the card’s removal).

§4.4 The 156-pair inter-root verdict matrix (G18)

The inter-root directed-pair verdict matrix is the deliverable of Phase 5b T12 (analysis/inter-root-causal-graph/directed-pairs-{Y1,Y2}.json). For each of the 156 ordered (chain_i → chain_j) pairs, a verdict is assigned: PROBABLE_CAUSE (strong PMID-anchored mechanism + direct LC-cohort observation + intervention or natural-experiment corroboration), PLAUSIBLE (mechanism PMID-anchored but LC-cohort observation or corroboration thinner), INSUFFICIENT_EVIDENCE (mechanism plausible but not verdicted at this corpus depth), EQUIPRIMORDIAL (both directions equally supported; no causal asymmetry detectable), or RULED_OUT (mechanism positively refuted by available evidence).

The combined Y1 + Y2 distribution across the 156 pairs is:

Verdict	Count	Fraction
PROBABLE_CAUSE	33	21.2 %
PLAUSIBLE	58	37.2 %
INSUFFICIENT_EVIDENCE	56	35.9 %
EQUIPRIMORDIAL	9	5.8 %
RULED_OUT	0	0.0 %

The matrix is rendered as exhibit G18 with chains on both axes and cells colored by verdict tier. No pair is RULED_OUT — meaning no candidate inter-chain causal claim is positively refuted by the available evidence — but only 21 % are PROBABLE_CAUSE, which is a measured rather than over-claimed picture. The 56 INSUFFICIENT cells (36 %) constitute a substantial future-research priority list catalogued in directed-pairs-{Y1,Y2}.json’s insufficient_pairs field; these are the pairs where the chain biology suggests a directional causal relationship but the corpus does not yet support a verdict.

The R-chain out-degree distribution (R1, R2, R3 outgoing PROBABLE_CAUSE counts) is the empirical signal that drives the three-root architectural claim (§4.1, CF-1). R1 has the most outgoing PROBABLE_CAUSE arrows (7 of 12 possible); R2 has 2 (R2 → R3, R2 → H10); R3 has 4 (R3 → H6, H7, H10, H13). The H-chains all have zero outgoing PROBABLE_CAUSE arrows into R-chains, which is the asymmetry that makes the three-root model defensible.

§4.5 The H7 ↔︎ H12 cycle and four asymmetric near-cycles (G19)

The 156-pair analysis surfaces one true bidirectional PROBABLE_CAUSE cycle: H7 ↔︎ H12 (autonomic ↔︎ HPA), rendered as exhibit G19 and documented in cycles.md. The forward direction (H7 → H12) is anchored by the observation that chronic sympathetic-parasympathetic dysregulation in POTS-like LC phenotypes drives sustained corticotropin-releasing hormone (CRH) release, which habituates the hypothalamic-pituitary-adrenal axis and progressively exhausts adrenal cortisol responsiveness; the phenotypic signature is the recurring finding of hypocortisolism in established autonomic-dysfunction-dominant LC subsets (Klein 2023 MY-LC PMID 37748514). The reverse direction (H12 → H7) is anchored by the observation that hypocortisolism reduces sympathetic vascular responsiveness — the Addison phenocopy — and amplifies orthostatic intolerance; the reverse direction is therapeutically supported (cortisol replacement is a recognized POTS therapy in hypocortisolemic patients).

Four asymmetric near-cycles surround H7 ↔︎ H12, each with a strongly supported PROBABLE_CAUSE dominant flow and a mechanistically plausible but evidentially thinner PLAUSIBLE reverse direction:

• NC01 H10 → H7 (with H7 → H10 PLAUSIBLE). Dominant: brainstem neuroinflammation (NTS, DVC, RVM) drives autonomic dysfunction; TSPO-PET brainstem microglial activation (Braga 2023 PMID 36571857) is the imaging anchor. Reverse: vagal afferent signaling modulates brainstem microglial state.
• NC02 H8 → H10 (with H10 → H8 PLAUSIBLE). Dominant: mast-cell tryptase activates microglial PAR2; mast-cell histamine activates microglial H1R/H4R (Zhang 2012 PMID 22531394, extended to LC by Theoharides 2017 PMID 28987303). Reverse: microglial cytokines (IL-6, TNF-α, CCL2) sensitize meningeal and dural mast cells in chronic phase.
• NC03 H10 → H12 (with H12 → H10 PLAUSIBLE). Dominant: chronic central neuroinflammation drives HPA exhaustion via hypothalamic inflammation and circulating-cytokine effects on adrenal cortisol-response capacity (Silverman & Sternberg 2012 PMID 22394319; Klein 2023 LC hypocortisolism cohort PMID 37748514). Reverse: glucocorticoid signaling is a tonic brake on microglial inflammatory activation; hypocortisolism removes the brake.
• NC04 H7 → H8 (with H8 → H7 PLAUSIBLE). Dominant: sympathetic catecholamines and sensory neuropeptides (substance P, CGRP) drive mast-cell degranulation; Theoharides 2017 PMID 28987303 is the canonical mechanistic anchor and the POTS-MCAS overlap in LC is the phenotypic integration. Reverse: mast-cell mediators modulate autonomic afferents and amplify sympathetic tone.

Two further asymmetric pairs surface in the directed-pair data scan: H5 → H10 with H10 → H5 PLAUSIBLE (autoantibodies driving CNS neuroinflammation, with reverse CNS-mediated B-cell modulation; synthesis §6 treats this as the fifth near-cycle on the chronic-tail surface) and H6 → H10 with H10 → H6 PLAUSIBLE (mitochondrial DAMPs driving microglial activation; synthesis treats this as a hierarchical edge rather than a near-cycle). Tasoula 2026 strengthens the H6 → H10 mechanistic edge at the molecular level: hamster cortex (mPFC) shows persistent OXPHOS repression at 31 dpi while sensory and deep-brain regions display heterogeneous metabolic-immune trajectories, and human frontal-cortex autopsy samples exhibit OXPHOS inhibition + HIF-1α + RAAS activation + immune upregulation — a region-specific brain mitochondrial-stress pattern that supplies a peer-reviewed multi-omics substrate for the cortical neuroinflammation signature H10 anchors (Greene 2024 PMID 38388736).

The cycle layer concentrates in the H7/H8/H10/H12 central-amplifier cluster. This is structurally consistent with the H12 reclassification (§4.6, CF-7) and with the H11-biology vindication (§4.7, CF-10). The atlas is not a strict DAG; it is a partially-cyclic graph with one true cycle and four (or six, on the inclusive count) asymmetric near-cycles. The cycle annotation layer is the manuscript’s transparent representation of what the T1 Bayesian noisy-OR DAG cannot natively encode (the T1 author’s own audit memo §5.3 acknowledges this limitation: “Real biology has feedback loops (BBB ↔︎ microglia ↔︎ IL6 ↔︎ TNFα) that this DAG can’t represent”).

§4.6 H12 reclassification (CF-7)

H12 (HPA axis) is reclassified from “parallel root” (its v3.1 classification) to “central downstream amplification hub” (CF-7). The empirical support is convergent across three techniques:

1. T12 inter-root verdicts (Y1 + Y2 combined): H12 has 3 PROBABLE_CAUSE incoming arrows (R1 → H12 in Y1; H7 → H12 and H10 → H12 in Y2) and 3 PLAUSIBLE incoming arrows (H5, H6, H8 → H12). The Y2-side outgoing H12 → H7 is PROBABLE_CAUSE (the reverse of the C01 cycle); H12 → H8 and H12 → H10 are PLAUSIBLE outgoing.
2. T1 Bayesian network inter-root posteriors: H7 ↔︎ H12 and H10 ↔︎ H12 posterior weights are elevated; R1 → H12 posterior is mid-range. H12 sits in the middle (not the upstream) layer of the MCMC inter-root posterior heatmap.
3. T8 cross-domain validation: H12 is VALIDATED against the Sheehan / autoimmune hypopituitarism comparator (PMID 18230820) as a downstream-of-autoimmune mechanism, not as a primary causal driver in those comparator syndromes.

The reclassification confidence is [CONSISTENT] not [ESTABLISHED] under the atlas pre-reg §7.1 vocabulary, per the honest-gap policy and the explicit T11 disagreement D-8: the Captain’s-brief threshold of ≥ 4 PROBABLE_CAUSE incoming is not formally crossed (3 observed). Phase 7 disposition PR-2 confirms this: the formal threshold is one short, no atlas mutation is applied, and Phase 9 manuscript handles the narrative re-framing. The chain text is updated to describe H12 as a central downstream amplifier with a bidirectional reinforcement loop to H7; the v3.3 corpus refresh + Phase 5b rerun on H12 incoming is the trigger event for upgrading to [ESTABLISHED].

The reclassification has clinical implications: H12 is no longer presented as a parallel root from which independent therapeutic intervention can target an upstream node (e.g., glucocorticoid replacement as a primary LC treatment), but as a downstream hub whose hypocortisolism phenotype reflects upstream R1, H7, H10 driving forces. Intervention at the H12 layer (e.g., fludrocortisone in aldosterone-stratified LC-POTS subsets) is symptomatic rather than disease-modifying in the hierarchical model. This is the correct framing for the v3.3 priority IM-4 (fludrocortisone-vs-placebo LC-POTS RCT; §12.2).

§4.7 The H11 biology vs framework distinction (CF-10)

H11 is the lowest-confidence chain in the atlas (T5 vulnerability 3.0/10) and the chain whose mere inclusion is structurally meta-recursive: it is the chain that red-teams single-driver hypotheses, included so the atlas itself can red-team it. The Phase 6 T11 capstone synthesis resolved the H11 question with a deliberate two-frame distinction (CF-10, v3.2-phase6-synthesis.md §6), which the Phase 7 strict re-tier audit subsequently made machine-verifiable (phase7-h11-biology-vs-framework-split.md).

H11-as-paper. The Groysman 2026 Frontiers in Hypothesis & Theory publication that the H11 chain wraps fails the atlas’s C1–C8 rubric: C1 (temporal precedence) is self-disclaimed by the framework; no PRISMA, PROSPERO, or GRADE methodology; single-author provenance in a Hypothesis-and-Theory venue; ~50 bidirectional edges asserted without per-edge mechanism, per-edge weight, or per-edge falsifier. T6 PRISMA scores H11 at 9.0/27 (lowest of 13 chains); T5 vulnerability scores H11 at 3.0/10; T2 MR has no instrument for H11 (no manipulable biological node); T9 counterfactual identification finds zero clean do() interventions on H11. T7 NMA flagged H11 as falsified by the temporal asymmetry of acute-Paxlovid (HR ≈ 0.74) vs LC-Paxlovid (four sequential nulls): a flat-equal-marginal-effect prediction is structurally incompatible with one node carrying all replicated positive RCT signal and the same molecular target’s temporal sibling carrying replicated null RCT signal. Phase 7 demoted all five H11 framework cards (H11.L02–L05) to SPECULATION at 1–2/8 scores.

H11-as-biological-claim. The underlying assertion that LC is multi-mechanism with feedback loops among mediators is empirically supported by four independent Phase 5/5b stress tests:

1. T12 cycle detection (Phase 5b Y2): a true bidirectional PROBABLE_CAUSE cycle exists at H7 ↔︎ H12, plus four asymmetric near-cycles (§4.5).
2. T7 NMA topology: empirical signature is node-dominated and temporally-anchored hierarchical, not flat (R1-acute carries 3 replicated positive RCTs; R1-LC carries 4 replicated nulls; SIM01 carries 1 positive; everything else uncontrolled or null). The H11 strong-form flat prediction is empirically refuted.
3. T8 cross-domain ME/CFS overlap: 8/13 chains validate or partial-validate against ME/CFS; LC’s mechanism stack (R2 dysbiosis, H2 T-cell exhaustion, H3 IFN-I drive, H6 mitochondrial, H10 neuroinflammation, H12 HPA) is concentrated in shared mediator nodes consistent with multi-mechanism final-common-pathway. The biological framing of H11 — that interactions among mediators are real — is empirically continuous with ME/CFS.
4. T10 cross-syndrome NC signature: 7/13 chains partial-fail cross-syndrome NCE on the 12-month chronic-tail phenotype. The signature is consistent with shared mediator pathways across post-infectious / post-inflammatory states (§9).

The hierarchical-DAG-with-annotated-cycles absorbs both views. R1 is the principal upstream initiator (CF-9). R1, R2, R3 are the three structural roots (CF-1). The H7 ↔︎ H12 cycle is real and annotated (§4.5). The asymmetric near-cycles H10 ↔︎ H7, H8 ↔︎ H10, H10 ↔︎ H12, H5 ↔︎ H10, H7 ↔︎ H8 are annotated with dominant-flow direction. The atlas claims a hierarchical initiation structure (R1-dominated, temporally-anchored, prevention-frame-asymmetric vs treatment-frame) and a partially-flat maintenance structure (mediator-feedback-loop-rich, cross-syndrome-shared). This is the empirically defensible reading.

We do not claim H11 is biologically wrong. We claim that the Groysman 2026 framework is methodologically and evidentially weak (C1–C8 fail; Phase 7 SPECULATION for all framework cards) and that its biological insight is subsumed by the hierarchical-DAG-with-annotated-cycles model, which makes stronger and more falsifiable predictions. The atlas’s #1 chain (R1) is presented as most-supported but not sole driver; the inter-root cycle analysis explicitly engages multi-driver causality. The dedicated H11 talking-point block (§2.4) extends this distinction for reviewers.

§4.8 The 6 LC-specific cells (cross-domain matrix)

T8 cross-domain coherence (analysis/cross-domain/per-chain-per-domain-matrix.json) tests each of the 13 chains against 5 comparator literatures (ME/CFS, POTS, PTLDS, post-Ebola, autoimmune syndromes) for 65 chain-comparator cells. The verdict distribution is:

Verdict	Count	Fraction
VALIDATED	19	29.2 %
PARTIAL	19	29.2 %
INSUFFICIENT	21	32.3 %
LC_SPECIFIC	6	9.2 %

The 6 LC_SPECIFIC cells (the only chain-comparator cells where the LC mechanism is not paralleled in the comparator literature) cluster in two places: R1 sustained S1 antigenemia + spike-protein persistence at 12+ months post-infection (Stein PMID 36517603, Goh PMID 34248921, Swank S1 longitudinal proteomics) — paralleled in no comparator syndrome — and H11 meta-claim (the flat-network framework structure is asserted only in LC literature). Every other chain has at least one comparator literature with VALIDATED-or-PARTIAL coherence.

This is the empirical signature behind two key findings: (1) R1 antigenemia is the principal LC-distinctive feature (§9); (2) the chronic-tail mechanism stack (R2, H2, H5, H6, H7, H8, H10, H12) is shared with the post-viral / autoimmune-syndrome spectrum — the cross-syndrome non-specificity that constitutes the atlas’s structural empirical liability (CF-3, §9). The hierarchical model handles this picture cleanly: the LC-distinctive initiation (sustained R1 antigen) feeds into the cross-syndrome-shared chronic maintenance machinery.

§4.9 Sex-typed early-vs-late disease evolution — the H6 inversion (CF-11)

A structural finding that no prior LC review has surfaced emerges when the AI pipeline scores per-card sex-hormone modulation (estrogen, progesterone, androgen) across the full 222-card corpus. Every link card in the v3.2 atlas carries a sex_modulation schema field (§3.3) populated by the AI curator with per-hormone effect direction (amplifies / suppresses / neutral / unknown) plus a rationale paragraph and supporting PMID array. The corpus-scale distribution is striking and was not visible to chain-level review.

The dominant signature: estrogen-amplifies + androgen-suppresses, 158 of 222 cards (71.2 %). Across nine of the thirteen chains — R1, R2, H2, H3, H5, H7, H8, H10, H11 — the modal pattern is that estrogen amplifies the mechanism and androgen suppresses it. The biological substrate is varied (estrogen modulation of B-cell class-switching for the autoantibody chains H2/H5, estrogen modulation of mast-cell histamine release for H8, estrogen modulation of microglial reactivity for H10/H11, estrogen modulation of sympathetic–parasympathetic balance for H7), but the directional pattern is consistent: female-physiological hormone milieu amplifies the chronic-tail amplifier stack. This is the structural mechanism behind the ~2:1 female-skewed LC prevalence reported across the international cohorts and the well-documented exacerbation of LC symptomatology in the luteal phase and during perimenopause (Davis 2023; Klein 2023 MY-LC PMID 37748514; PMID 35933347; PMID 35912863).

The H6 inversion: estrogen-suppresses + androgen-neutral, 17 of 20 H6 cards (85 %). The mitochondrial cell-danger-response chain is the sole chain in the atlas that inverts the dominant sex-typing pattern. 17 of 20 H6 cards encode estrogen as suppressing the mitochondrial CDR / PGC-1α / sirtuin / SCFA-energetic axis (estrogen is generally pro-mitochondrial-biogenesis via ERα → PGC-1α → NRF1/NRF2 → mtDNA replication; the H6 cards report this protective effect as inverse with respect to the chain’s pathological direction), and only 2 of 20 H6 cards encode androgen as suppressing. The H6 inversion is anchored on the Appelman / Wüst 2024 PMID 38381103 activity-matched CPET-PEM cohort (the only chain to pass the T10 cross-syndrome negative-control battery outright; §5.7), the Naviaux 2016 metabolomic CDR signature (cross-domain VALIDATED in ME/CFS; §5.6), Tasoula 2026 Front Immunol multi-omics demonstration of sustained mtDNA-encoded OXPHOS transcript suppression in PBMCs out to 12 months post-infection (acute disease severity stratification rather than symptom stratification — an important honesty caveat the chain rubric inherits, but the molecular substrate is peer-reviewed and longitudinal), and a series of LC-cohort mitochondrial-function papers with sex-stratified analyses. The H6 inversion is the structural reason that male LC patients are disproportionately represented in the late-stage mitochondrial / energetic-failure / PEM-dominant LC phenotype.

The implication: a sex-typed early-vs-late disease evolution. The combined corpus-scale pattern resolves into a structural prediction the field has not yet operationalized:

• Female LC patients are disproportionately represented in the early autoimmune / autonomic / mast-cell / neuroinflammatory amplification phase of LC (chains H5/H7/H8/H10/H12). Estrogen amplifies the chronic-tail amplifier stack; female-physiological hormone milieu drives faster and more severe symptom emergence at 0–6 months post-acute. The female-LC clinical phenotype is dysautonomia-dominant, mast-cell-dominant, autoimmune-driven, with hormone-cycle-linked flares.
• Male LC patients are disproportionately represented in the late-stage mitochondrial / energetic-failure / PEM-dominant phase of LC. The androgen-neutral H6 profile combined with the absence of estrogen-amplification of the upstream amplifier chains means male LC tends to emerge more slowly, with a chronic-tail trajectory weighted toward mitochondrial CDR, PEM, and metabolomic CDR signatures. The male-LC clinical phenotype is energy-failure-dominant and post-exertional-malaise-dominant.

This sex-typed early-vs-late disease-evolution mechanism is CF-11 in the capstone-finding inventory and is one of the manuscript’s three principal AI-enabled biological discoveries (alongside the 156-pair directed-pair verdict matrix in §4.4 and the H7↔︎H12 self-sustaining maintenance cycle in §4.5). It is not derivable from any chain-level review because it requires the corpus-scale aggregation of per-card sex-modulation scoring, and it directly motivates three v3.3 research-agenda priorities (§12): (i) sex-stratified trial designs as the default for all LC RCTs targeting the chronic-tail amplifier chains; (ii) time-stratified enrollment (0–6 months for the female-amplifier phase, 12–+ months for the male-mitochondrial phase) to detect effect sizes the underlying biology supports; and (iii) a dedicated peri-/post-menopausal LC cohort study to characterize the hormone-fluctuation phase of female LC at biomarker depth. The current LC trial literature pools sex and time post-acute by default and is therefore structurally underpowered to detect the sex-typed mechanistic dichotomy the v3.2 atlas surfaces.

The per-card sex-modulation JSON arrays are at analysis/atlas/v3.2.1-pathways/<chain>/<card>.json and are queryable via analysis/sex-modulation/per-chain-summary.json (corpus-scale aggregation script published in supplementary materials). A reviewer who wishes to audit the H6-inversion claim can query the 20 H6 cards directly and inspect the estrogen-rationale + supporting-PMID array per card; the dominant 71.2 % estrogen-amplifies + androgen-suppresses pattern across the other twelve chains is similarly queryable.

§4.10 The primacy of the three causal roots (CF-1 + CF-9 — the empirically-established headline)

The three root causal chains are visualized in detail in exhibits G40 (R1 viral persistence), G41 (R2 dysbiosis with SIM01 keystone), and G42 (R3 endothelial / microclot with CN-3 replication callout). Each diagram traces the chain from initiator → mechanism nodes → mediator outputs → symptom domains, with anchor PMIDs and tier badges. The ten amplification hubs are detailed in exhibits G43 (H2), G44 (H3), G45 (H5), G46 (H6 with sex-inversion callout), G47 (H7 cycle partner), G48 (H8), G49 (H10 SPOF), G50 (H11 two-frame synthesis), G51 (H12 cycle partner), and G52 (H13 pediatric).

The single most consequential structural result of the v3.2 atlas is the empirically-established primacy ordering of the three causal roots of Long COVID. Prior LC syntheses have asserted root-and-amplifier structures qualitatively; the v3.2 atlas derives the primacy ordering empirically from the 156-pair directed-pair verdict matrix (§4.4, T12), the Bayesian-network composite-influence ranking (§3.7, T1), the Cinelli-Hazlett sensitivity bounds (§5.4, T3), and the cross-domain coherence matrix (§5.6, T8). The four-technique structural convergence is itself the empirical defense of the primacy ordering; any one of the four techniques is rebuttable in isolation, but the convergence is not.

Root 1 — R1: sustained SARS-CoV-2 viral / antigen persistence — is the singular upstream initiator of Long COVID. Four convergent empirical signatures establish R1 as the primary cause of LC at the level the corpus supports.

• T12 directional reach. R1 produces 33 PROBABLE_CAUSE outgoing edges into downstream chains in the 156-pair verdict matrix — more than any other chain in the atlas, and more than R2 and R3 combined at the PROBABLE_CAUSE tier. No incoming PROBABLE_CAUSE edge from any other chain reaches R1; R1 is structurally upstream in the empirical directed-pair scoring.
• T8 cross-domain LC-specificity. R1 is the only chain whose chronic phenotype is LC_SPECIFIC in the 65-cell cross-domain coherence matrix at the 12+ month timepoint. The sustained S1 antigenemia signature (Swank 2024 PMID 39389851 multicohort), the autopsy multi-organ persistence (Stein 2022 PMID 36517603), the gut-mucosal persistence (Goh 2022 PMID 36131932 and Cheung 2024 follow-up), and the reservoir-review framework (Proal & VanElzakker 2023 PMID 37667052) collectively establish R1 as the only chain not paralleled in any comparator post-infectious / chronic-illness literature (ME/CFS, POTS, PTLDS, post-Ebola, autoimmune syndromes). Every other chain has at least one comparator literature with VALIDATED-or-PARTIAL coherence. R1 stands alone.
• T3 Cinelli-Hazlett robustness. The R1 backbone survives Cinelli-Hazlett sensitivity perturbation at the highest robustness percentile in the atlas. The R-value threshold to overturn any R1 PROBABLE_CAUSE edge exceeds the conservative threshold by a substantial margin.
• T9 Pearl-identifiability — honest caveat. R1’s therapeutic identifiability is a separate question from R1’s causal primacy. R1-targeting therapeutics (nirmatrelvir during established LC, anti-spike passive transfer, anti-S1 monoclonals) fail Pearl identifiability at the strict-rubric threshold for use against established LC because the downstream maintenance cycle (§4.5) is self-sustaining at 12+ months even if R1 antigen reservoir is reduced. R1 is the causal-primacy headline; R1 is not the therapeutic-priority headline. The Paxlovid prevention-vs-treatment asymmetry (Xie 2023 PMID 36951829 prevention positive; Geng 2024 PMID 38848477 STOP-PASC treatment null) is the empirical signature of this distinction.

Root 2 — R2: gut microbiome dysbiosis — is the second-ranked upstream root and the source of the sole Pearl-defensible therapeutic claim. R2’s empirical signatures complement R1’s: where R1 dominates by directional reach and LC-specificity, R2 dominates by mediator-cluster influence and therapeutic identifiability.

• T1 Bayesian-network composite influence. X.Dysbiosis ranks #1 by Bayesian-network composite influence across the 27-mediator polytope (MCMC composite influence 0.234; the next-ranked mediator X.MitoCDR scores 0.186, X.Autoantibody scores 0.171). The composite-influence metric integrates per-edge posterior probability, downstream symptom-polytope coverage, and per-mediator Cinelli-Hazlett robustness. R2’s upstream node is the most-influential single mediator in the atlas by this composite measure.
• T9 Pearl identifiability. SIM01 → X.Dysbiosis is the sole Pearl-defensible therapeutic claim the atlas produces (composite priority score P = 9.00 vs 1.96 for the runner-up; §7; finding #4). This is not coincidence: R2 is the upstream root whose identifiability geometry (back-door / front-door admissibility under the no-unmeasured-confounder assumption) permits Pearl identification under the v3.2 mediator DAG.
• T8 cross-domain. R2 backbone is VALIDATED in ME/CFS at the T8 layer (the IBS-ME/CFS overlap literature carries an R2-equivalent backbone), supporting R2 as a genuine causal upstream node rather than a comorbid downstream signature. R2 partially-fails the cross-syndrome NCE (7/13 chains do) but the partial-failure pattern is the cross-syndrome-shared-mediator signature, not an R2-specific weakness.
• T2 Mendelian randomization. R2 has no clean cis-eQTL instrument and is therefore MR-untestable; this is the principal honest caveat to R2’s causal primacy claim. The MR-null on R2 is a methodological limit, not an empirical refutation (§5.4).

Root 3 — R3: endothelial dysfunction / microvascular coagulopathy — is empirically established as a structurally-distinct third root, not collapsible to R1 or H10. The R3 partition decision is the most-contested structural claim in the atlas, and it survives the multi-technique empirical stress-tests.

• T12 directional structure. The directed-pair matrix returns R3’s strongest outgoing reach into the chronic-tail mediator chains (H6, H7, H10) rather than into R1 or H10 alone. R3 → H10 carries the strongest single edge, but R3 → H6 and R3 → H7 are also PROBABLE_CAUSE. The cycle-graph analysis (§4.5) confirms R3 as upstream-not-downstream of H10: H10 → R3 is INSUFFICIENT, while R3 → H10 is PROBABLE_CAUSE.
• T2 Mendelian randomization — the ABO signal. The Lammi 2025 GWAS PMID 40399555 identifies a genome-wide-significant ABO locus signal for LC outcome. The R3 partition decision predicts that the ABO signal should be R3-mediated (microvascular / blood-group) rather than R1-mediated (antigen-persistence). The MR-instrument geometry supports the R3-mediated interpretation. This is the strongest single piece of MR evidence the atlas integrates.
• T3 Cinelli-Hazlett robustness. The R3 microclot backbone survives Cinelli-Hazlett sensitivity perturbation at the CONSISTENT robustness percentile (Pretorius 2021 PMID 34425843 anchor; Kell & Pretorius 2024 review; Robson 2024 follow-up replication).
• Honest caveat. R3 carries CONSISTENT (not ESTABLISHED) tier in v3.2.1 because the sham-controlled HELP-apheresis RCT does not yet exist. A positive readout would promote R3 to ESTABLISHED; a null readout would demote R3-microclot-as-causal to EMERGING. The R3 partition decision is empirically defensible at CONSISTENT tier and is correctly tier-marked. R3’s primacy as a third root is established at the structural level; R3’s therapeutic actionability awaits the sham-RCT readout.

Together, R1 + R2 + R3 constitute the empirically-established three-root upstream cause of Long COVID. No other chain in the atlas carries upstream-root status under the multi-technique convergence. The H-chains (H2 through H13) are downstream amplifiers in the hierarchical structure (§4.1), with the chronic-tail amplifier cluster (H6, H7, H8, H10, H12) carrying the self-sustaining maintenance machinery (§4.5). The hierarchical-DAG-with-annotated-cycles architecture is the empirically-correct framing of the LC mechanism field, and the three-root primacy ordering is the central biological finding of v3.2 alongside the sex-typed early-vs-late disease evolution (§4.9), the H7↔︎H12 self-sustaining maintenance cycle (§4.5), the SIM01 Pearl-defensible therapeutic claim (§7), and the 4-universal-fail field diagnosis (§8).

The primacy ordering directly motivates the v3.3 trial-design priority queue (§12): R1-targeting prevention trials (acute-COVID nirmatrelvir + S1-antigenemia-stratified follow-up); R2-targeting established-LC trials (SIM01 multi-center replication + BHC-202 + butyrate supplementation arms); and R3-targeting sham-controlled HELP-apheresis as the definitive R3 actionability test. The trial-design queue is the operational consequence of the primacy ordering, and the field’s current trial portfolio is structurally mis-aligned with the primacy ordering — the bulk of LC trial-funding has historically gone to downstream-amplifier targets (IVIG-SFN, anti-mast-cell, beta-blockade) rather than to the upstream-root primacy ordering the v3.2 atlas now establishes.

---

§5 Evidence quality and sensitivity

§5.1 Bayesian network (T1)

The v3.2 Bayesian network is constructed as a 482-edge noisy-OR DAG over the 13 chains, 27 mediator nodes, and 117 polytope cells (analysis/bayesian-net/edge-weights.json). Two complementary inference runs were executed against the same edge weights: (a) a PyMC NUTS MCMC sampler with 6 chains × 2,000 warmup × 4,000 sampling iterations, and (b) an analytic Beta-Binomial conjugate run that exploits the noisy-OR likelihood’s conjugacy with Beta-distributed edge-weight priors. Both runs converge on the same qualitative no-bottleneck conclusion (CF-2) and on the same posterior topology (exhibit G23).

MCMC convergence diagnostics are reported in the run log: R-hat = 1.0000 (Gelman-Rubin diagnostic; perfect convergence threshold ≤ 1.01), ESS_min = 6,442 (effective sample size minimum across all parameters; convergence threshold typically ≥ 400), 0 divergences (HMC divergent-transition count; healthy runs target 0; the atlas’s run is geometrically benign). The posterior predictive check yields r = 0.984 between observed and posterior-predicted polytope-cell occupancy (exhibit G24). The T1 author’s audit memo §5 explicitly notes that the Beta-Binomial conjugacy yields automatic PPC accuracy and that r = 0.984 is expected given the prior structure, not evidence of generative correctness — this honest caveat is preserved in the disagreement table (D-7) and propagated to §11.3.

No single mediator is bottleneck-critical (CF-2). Across both runs, the maximum mediator knockout impact on strongest-path probability (averaged over 12 roots × 9 symptoms after removing each mediator one at a time and recomputing the posterior strongest-path probability) is ≤ 0.012. The top-ranked mediators by knockout (X.Tcell_exhaustion 0.0112, X.SympNS 0.0108, X.Endothelium 0.0064 in the analytic run) are strongest single-source bridges — their removal is partially absorbed by parallel routes (e.g., NETs → microclot, TNFα → endothelium). The composite-influence MCMC ranking (X.Dysbiosis 0.234, X.SCFA 0.114, X.Histamine 0.061, X.Tcell_exhaustion 0.049, X.Complement 0.048) and the analytic-knockout ranking are not in tension; they answer different questions (D-3, §3.7). The atlas reports both: composite influence for mediator-hub argumentation (§4.2, §7); knockout impact for intervention-prioritization (§6.2).

The robustness under two methodological choices (analytic vs MCMC), two sensitivity metrics (composite vs knockout), and two topologies (482-edge canonical vs 854-edge expanded) supports CF-2 at the [ESTABLISHED] confidence tier.

§5.2 Cinelli-Hazlett robustness (T3)

The Cinelli-Hazlett robustness value RV is a closed-form bound on the strength of an unobserved confounder required to overturn an observed estimate. For a card with regression statistic t and degrees-of-freedom df: RV = 0.5 × (√(f⁴ + 4f²) − f²) where f² = t²/df. A card with RV > 0.30 is HIGH_ROBUST (robust to a confounder explaining > 30 % of residual variance on both treatment and outcome).

Of 58 ESTABLISHED-tier link cards in v3.2, 9 yield closed-form RV from primary-citation regression output. All 9 score HIGH_ROBUST. The top 5 are:

Rank	Card	RV	Chain
1	H8.L03	0.772	Mast-cell tryptase / chymase elevation in LC vs convalescent controls
2	H13.L02	0.669	Pediatric inflammatory cytokine panel + adipokine modulation
3	H10.L01	0.648	Acute SARS-CoV-2 → CNS-relevant antigen exposure
4	H6.L07	0.591	CDR → skeletal-muscle mitochondrial myopathy
5	H10.L07	0.529	CNS cytokines → neuroaxonal injury (serum NfL)

For the remaining 49 ESTABLISHED-tier cards, a structured heuristic RV was applied (cited papers’ sample size, effect direction consistency, and replication-group count combined into a per-card 0–1 score). 14 cards score HIGH_ROBUST (RV > 0.30) under heuristic; 29 score MODERATE (0.10–0.30); 13 score LOW or FRAGILE (< 0.10). The 13 LOW_ROBUST / FRAGILE cards concentrate in R1.L05 and R1.L06 (cytokine-cascade in-vitro mechanism) and H11 framework-evaluation cards. The full distribution is rendered as exhibit G25 and reported in the cinelli-hazlett-summary.md companion file.

Headline interpretation: every ESTABLISHED claim that is closed-form RV-computable survives a confounder explaining > 30 % of residual variance on both treatment and outcome. The structurally honest caveat (§11.5): only 9 of 58 (16 %) of ESTABLISHED cards are closed-form computable; the bulk inherit the heuristic RV, which provides a defensible relative ranking but not absolute calibration.

§5.3 Per-link polytope leverage and SPOFs (T4)

The per-link polytope-leverage analysis (analysis/sensitivity/per-link-polytope-loss.json) classifies each of the 222 link cards by its single-edge-cut impact on the 117-cell polytope. A card whose removal drops the polytope coverage for ≥ 1 cell is a single point of failure (SPOF). After the Phase 6 correction (removal of H11.L20, a polytope-symptom-convergence summary card that is not a manipulable biological node — exhibit G26), the atlas has 4 SPOFs of 222 cards = 1.8 %:

Rank	Card	Chain	Tier (v3.2 / v3.2.1)	Cells lost	Why it is a SPOF
1	H3.L10	H3 composite amplification	ESTABLISHED / CONSISTENT-pending-falsifier (PR-1)	1 (H3 / sleep)	Composite cytokine + autoantibody + BBB disruption + microglial activation load → amplified brain fog + sleep terminal
2	H10.L18	H10 neuroinflammation	CONSISTENT / EMERGING	1 (H10 / SFN)	DRG + satellite glial cell neuroinflammation → small-fiber neuropathy terminal
3	H11.L13	H11 musculoskeletal	CONSISTENT / EMERGING	1 (H11 / GI)	Acute SARS-CoV-2 + vagus neurotropism → vagus nerve inflammation
4	H11.L16	H11 musculoskeletal	CONSISTENT / EMERGING	1 (H11 / SFN)	Endothelial glycocalyx shedding + microclot + senescence → capillary rarefaction / perfusion heterogeneity in skeletal muscle

109 of 192 inter-card edges carry zero polytope leverage (parallel-path redundancy or non-symptom termination). The maximum single-edge cut by any card on the polytope is 8 cells (R3.L01, the first link in chain R3, reflecting the chain’s linear authoring — the cut is graph-topological, not biological-bottleneck). The polytope coverage is redundant by design: mean 5+ reaching paths per cell; 34 of 117 cells (29 %) robust with ≥ 5 paths.

The SPOFs are graph-theoretic, not biological-importance rankings. Each of the four cards has CONSISTENT or ESTABLISHED upstream support (v3.2 tier); their SPOF status means that the atlas’s polytope-symptom coverage in that specific cell currently depends on this single link card. Removal would not falsify the broader chain, but it would mark a polytope cell as uncovered, which is itself an honest finding (one symptom-chain pairing rests on one anchor). H3.L10 in particular is the subject of disagreement D-5 (T1 BN counterfactual leverage +27.3 pp vs T9 Pearl identifiability “provisional and contested”); the disposition is Phase 7 PR-1 (re-evaluate H3.L10 tier when HEAL-LC NCT07597902 reads out, estimated 2027).

§5.4 Mendelian randomization (T2)

The T2 Mendelian randomization layer (analysis/mr/v3.2/) tests four instruments against five outcomes (analysis/mr/v3.2-mr-instruments.json, v3.2-mr-results.json):

Instruments. ABO B-allele rs8176719 (cis-eQTL for ABO; multi-pathway exposure spanning vWF and complement); IL6R rs2228145 + 2 LD-r² > 0.95 proxies that act as effectively a single Wald-ratio (IL-6 trans-signaling axis); HLA-region SNP cluster (immune-MHC); ACE2 eQTL rs2074192 + 2 supporting cis-eQTL (germline ACE2 vulnerability). F-statistic minimum across the four instruments: 27.6 — passes the conventional weak-instrument threshold of 10.

Outcome GWAS. Lammi 2025 Nature Genetics PMID 40399555 (6,450 LC cases, 1,093,000 controls). The 2024 preprint (PMID 38233424) is cited only as cross-check; the published Lammi 2025 outcome is used for headline β / SE computation.

Bonferroni-robust results (α = 7.7 × 10⁻⁴):

• IL6R → fatigue β ≈ +1.5 (single-instrument Wald, p ≈ 0.001–0.003). The only Bonferroni-robust positive across the full 65-pair grid. The result is consistent with prior MR evidence on IL-6 trans-signaling in chronic-fatigue traits but is fatigue-domain-specific in LC; brain fog, dysautonomia, and dyspnea outcomes are null.
• ACE2 → LC β ≈ 0 with bounded CI (informative null). The germline ACE2 vulnerability hypothesis is empirically downweighted; the chain card-level integration is that R1 antigen persistence depends more on host-pathogen interaction details than on germline ACE2 expression variance.
• ABO → LC β = +0.27, p = 0.027 (single B-allele SNP, weak nominal). Does not survive Bonferroni; consistent with the established ABO → severe-COVID association but insufficient for a microclot causal claim. The T2 audit memo notes that an earlier draft of the analysis framed this as P < 10⁻¹¹; the empirical β / p above is reproducible from the cached harmonised summary statistics with seed = 42, and the earlier figure does not appear in the audit memo. T11 disagreement D-2 is the formal correction.
• HLA-region → LC and HLA-region → fatigue: uninformative due to pleiotropy intercept p = 0.018 (MR-Egger pleiotropy test marginally significant; the HLA region’s multi-trait associations cannot be cleanly disentangled by current MR methodology). The HLA pleiotropy is an honest methodological dead-end.

The MR forest plot is rendered as exhibit G27. The headline picture is mostly null at Bonferroni: only one Bonferroni-robust positive (IL6R → fatigue), and a known protective signal (ApoE4 → LC, β = −0.140, p = 0.032 in IVW; MR-Egger and Weighted Median attenuate to null) which the Phase 3 redteam attributed most plausibly to survivorship bias (acute COVID excess mortality in autoimmune-vulnerable populations producing exactly this signal). The atlas does not interpret the ApoE4 protective signal as evidence against the H5 autoantibody chain.

LC lacks clean germline instruments. This is an honest result, not an analytical failure. The LC GWAS (Lammi 2025) is the largest published outcome cohort to date (6,450 cases), but Long COVID is a multi-mechanism syndrome and germline variants do not strongly select for any single mechanism. MR is a methodologically appropriate tool only for the small subset of LC mechanism claims that have clean cis-eQTL instruments (e.g., IL6R, ACE2); for the rest, the appropriate tool is the inter-root T12 directed-pair verdict (§4.4) and the T9 Pearl counterfactual (§6). The 64-of-65-pair null result is itself a structural finding consistent with CF-2 (§5.8): no single germline-instrumentable mediator has a population-scale knockout effect on aggregate LC outcome, and the one positive (IL6R → fatigue) is symptom-domain-specific rather than cross-cutting. The atlas treats this null-dominance as evidence for the multi-mechanism / parallel-redundancy architecture rather than as an indictment of any individual mechanism claim — and Phase-7’s strict rubric explicitly does not downweight a mechanism claim for failing an MR Bonferroni gate, because for most LC mechanism claims, MR is the wrong test.

§5.5 Chain vulnerability scoring (T5)

The T5 chain-vulnerability scoring (analysis/redteam/chain-vulnerability-scores.json) is the deliverable of an adversarial three-panelist simulated hostile peer review per chain. Each chain receives a 0–10 score, with the components: (a) ROBINS-I severity (40 %), (b) PRISMA compliance (15 %), (c) cross-domain coherence (15 %), (d) counter-narrative resilience (20 %), (e) pre-registered-falsifier discipline (10 %). The full per-chain rubric is in analysis/redteam/per-chain-vulnerabilities.md.

The per-chain distribution is reported in exhibit G28:

Chain	T5 score	Vulnerability class
H6	8.5	HIGH ROBUST
H10	8.0	HIGH ROBUST
R1	7.5	HIGH ROBUST
R2	7.0	ROBUST
H8	7.0	ROBUST
H13	6.5	ROBUST
H2	6.5	ROBUST
H3	6.0	MODERATE
R3	6.0	MODERATE
H7	6.0	MODERATE
H5	5.5	MODERATE
H12	5.5	MODERATE
H11	3.0	WEAK (by intentional design)

Mean across 13 chains = 6.5/10. Three chains score HIGH ROBUST (≥ 7.5): H6, H10, R1. Only H11 falls below 5.0, and H11’s status is meta-recursive by intentional design (the chain that red-teams the field’s single-driver hypotheses is itself rated weakest, which is methodologically appropriate and honestly disclosed). The H6 = 8.5 score reflects the lowest mean ROBINS-I severity in the atlas (1.25), the Appelman/Wüst 2024 PMID 38381103 activity-matched CPET-PEM design (T10 NC gold-standard), and H6’s strong cross-domain ME/CFS validation (PMID 32993883 mitochondrial dysfunction in ME/CFS).

The T5 vulnerability ranking is itself LLM-panelist-derived (§11.6). The ranking is sufficient for relative chain-strength comparison and for the §11.5 anticipated-reviewer-challenges framing; it is not a replacement for journal-grade external peer review.

§5.6 Cross-domain coherence (T8)

The T8 cross-domain coherence matrix (§4.8) tests each chain against 5 comparator literatures. The verdict distribution is 19 VALIDATED / 19 PARTIAL / 21 INSUFFICIENT / 6 LC_SPECIFIC; ME/CFS is the strongest comparator (8/13 chains VALIDATED-or-PARTIAL). The full 13 × 5 matrix is rendered as exhibit G29.

The strongest cross-domain anchors per chain are:

• R1 ↔︎ post-Ebola viral persistence (Varkey 2015 PMID 26465383) — VALIDATED for the upstream antigen-persistence-as-LC-driver claim.
• R2 ↔︎ ME/CFS gut dysbiosis (Giloteaux 2016 PMID 28095889; Xiong 2023 PMID 36450980) — VALIDATED for the X.Dysbiosis mediator-target pathway (this is the cross-domain anchor for SIM01, §6, §7).
• R3 ↔︎ autoimmune endothelitis — PARTIAL for endothelial dysfunction; LC-specific for the antigenemia-driven complement-coagulation cross-talk.
• H2 ↔︎ ME/CFS T-cell exhaustion (PMID 37452915) — PARTIAL; T-cell exhaustion is a shared post-viral final-common-pathway signature.
• H3 ↔︎ autoimmune IFN-I signature (Banchereau 2014 PMID 24566942) — PARTIAL; the SLE archetype IFN-I-driven inflammation is paralleled.
• H5 ↔︎ autoimmune DN2 extrafollicular response (PMID 30143188) — VALIDATED; the DN2 B-cell signature is independently documented in SLE.
• H6 ↔︎ ME/CFS mitochondrial (PMID 32993883) — VALIDATED; Appelman/Wüst PMID 38381103 is the gold-standard activity-matched CPET-PEM design. The T8 verdict is reinforced post-v3.2 closure by Tasoula 2026 Front Immunol doi:10.3389/fimmu.2026.1776555, which directly compares PCS-CFS and T2bFA-CFS vastus lateralis biopsies and finds shared core transcriptional programs (downregulated OXPHOS, elevated HIF signaling, mitochondrial stress) across the two CFS phenotypes — the first peer-reviewed same-tissue same-protocol molecular convergence test of the PCS↔︎ME/CFS mitochondrial overlap and the strongest single VALIDATED-tier reinforcement Tasoula contributes to the T8 matrix. The Phase 11f addendum (§14.10) does not re-tier the T8 matrix on this one paper alone, but H6 ↔︎ ME/CFS is moved from VALIDATED-by-comparator-citation to VALIDATED-by-direct-molecular-convergence within the existing tier band.
• H7 ↔︎ PTLDS small-fiber neuropathy (PMID 23390244) — PARTIAL; SFN is a post-infectious shared phenotype.
• H8 ↔︎ POTS-MCAS-EDS triad (PMID 33713596) — VALIDATED; the autonomic-MCAS overlap is independently documented.
• H10 ↔︎ post-sepsis TSPO-PET microglial activation (Greene 2024 PMID 38388736) — PARTIAL; the BBB / microglial neuroinflammation signature has post-sepsis parallel.
• H11 ↔︎ (meta-claim) — LC_SPECIFIC by construction (the flat-network framework is asserted only in LC literature).
• H12 ↔︎ autoimmune hypopituitarism / Sheehan (PMID 18230820) — PARTIAL; H12 HPA exhaustion paralleled in lymphocytic hypophysitis.
• H13 ↔︎ MIS-C cross-link to H5 + H10 — PARTIAL; pediatric immune-dysregulation phenotype paralleled.

The cross-domain matrix is the empirical foundation for the cross-syndrome non-specificity finding (§9). The chronic-tail mechanism stack is largely shared with ME/CFS, POTS, PTLDS, and autoimmune-syndrome spectrum; the LC-distinctive feature is the sustained R1 antigen-persistence node.

§5.7 Negative-control battery (T10)

The T10 negative-control battery (analysis/negative-controls/per-chain-negative-controls.json) applies the Lipsitch / Tchetgen-Tchetgen NCE / NCO / tracer framework to each of the 13 chains. The framework distinguishes:

• NCE (negative control exposure): an exposure that should not affect the outcome by the proposed mechanism. If the NCE association is non-null, the chain is suspect.
• NCO (negative control outcome): an outcome that should not be affected by the exposure. If the NCO association is non-null, the chain is suspect.
• Tracer: a within-chain biomarker that should track the mechanism. If the tracer does not track, the chain is suspect.

The per-chain verdict distribution is reported in exhibit G30:

Chain	NCE verdict	NCO verdict	Tracer verdict	Aggregate
R1	partial_fail (cross-syndrome 12-mo persistence)	n/a	passes (Stein autopsy, Goh biopsy, Swank S1)	partial_fail
R2	untested (PI-IBS NCE missing)	n/a	passes (Liu, Yeoh microbiome dynamics)	untested
R3	partial_fail (ThT layer cross-syndrome)	n/a	passes (Cervia-Hasler complement-coag)	partial_fail
H2	partial_fail (ME/CFS EBV reactivation)	n/a	passes (Klein 2023 dual signature)	partial_fail
H3	untested	n/a	passes (Klein 2023 IFN-I)	untested
H5	partial_fail (autoimmune DN2)	n/a	passes (Iwasaki passive transfer, Wallukat)	partial_fail
H6	passes (Appelman/Wüst CPET-PEM)	passes	passes	PASSES
H7	partial_fail (PTLDS SFN)	n/a	passes (Oaklander IENFD)	partial_fail
H8	untested	n/a	passes (Theoharides)	untested
H10	partial_fail (post-sepsis)	n/a	passes (Greene BBB-MRI, Visser TSPO-PET)	partial_fail
H11	structural NC paucity	n/a	n/a (framework, not biology)	structural
H12	partial_fail (Sheehan / autoimmune-hypopituitarism)	n/a	passes (Klein 2023 hypocortisolism)	partial_fail
H13	untested (growth-velocity NCO not run)	n/a	partial (Brodin, Buonsenso)	untested

Only H6 passes the negative-control battery outright. Seven chains (R1, R3, H2, H5, H7, H10, H12) partial-fail the cross-syndrome NCE — meaning the within-LC tracer evidence behaves as the chains predict, but the chronic 12-month phenotype is not yet differentiated from generic post-viral / chronic-illness patterns. Four chains (R2, H3, H8, H13) are untested at the NCE layer. H11 is structurally NC-paucity-bounded.

The cross-syndrome partial-fail signature is the empirical signature behind:

• CF-3 (cross-syndrome non-specificity is the atlas’s structural empirical liability; §9).
• CF-10 (the H11 biological claim — that LC is multi-mechanism with shared final-common-pathways — is partially vindicated; §4.7).
• The §9.4 implication: most “LC-specific” mechanisms are shared post-viral / chronic-illness final-common-pathways; the LC-distinctive feature is the sustained R1 antigenemia (Stein PMID 36517603, Goh PMID 34248921, Swank S1 longitudinal).

CF-6 disposition: vaccination is a positive control on (acute infection → LC), not an NCE against any chain. The Watanabe / Notaras / RECOVER vaccination literature documents that vaccination reduces acute COVID severity, which reduces LC incidence (a parallel to the Paxlovid acute-prevention HR ≈ 0.74 effect). This pattern is consistent with the atlas’s hierarchical causal structure but functions as a PC on the (acute → LC) arrow, not as an NCE. A true vaccination NCE would be an unrelated vaccine (e.g., influenza) → LC risk with null prediction; tested peripherally in EHR data and largely null, as expected.

§5.8 Triangulation: no single mediator bottleneck (CF-2)

The convergent finding from T1, T3, T4, and T8 is that no single mediator is bottleneck-critical. The maximum mediator knockout impact across both T1 inference runs is ≤ 0.012 on strongest-path probability — meaning that removing any single one of the 27 catalogued mediators reduces the downstream symptom-prediction probability by at most 1.2 percentage points. The top-5 most-robust Cinelli-Hazlett claims are distributed across four chains (H6, H8, H10, H13), not concentrated in one root. The four polytope-leverage SPOFs are 1.8 % of cards — and three of them (H3.L10, H10.L18, H11.L13) are framework-construction artifacts rather than biological chokepoints. The cross-domain coherence is concentrated in shared mediator pathways across post-viral syndromes (§5.6, §9). Capstone finding CF-2 is the formal statement of this convergence: there is no manipulable single-node bottleneck in the v3.2 atlas.

The structural and therapeutic implications of CF-2 are significant. First, LC is a multi-mechanism syndrome with parallel-path redundancy. The four-criterion convergence — T1 ≤ 1.2 % knockout impact, T3 distribution across four chains, T4 1.8 % SPOF prevalence, T8 cross-syndrome shared coherence — collectively rule out the single-driver framing that has dominated parts of the LC clinical-research discourse (e.g., “LC is primarily microclots,” “LC is primarily viral persistence,” “LC is primarily autoimmunity”). Each of those framings finds support in the atlas, but none survives sensitivity analysis as the exclusive driver. The atlas’s hierarchical-DAG-with-annotated-cycles architecture (§4.1) is the structural embodiment of CF-2: R1 is upstream and necessary, but downstream amplification through H6 / H10 / H12 is what produces the symptomatic phenotype, and these arcs are not collapsible to a single node.

Second, CF-2 has direct therapeutic implications. Single-target RCTs against single nodes — H7-only beta-blocker monotherapy for POTS, H5-only IVIG for autoantibody clearance, H6-only mitochondrial cofactor supplementation — should not be expected to deliver large effect sizes on aggregate LC symptom burden, even when they are mechanism-consistent. The CF-2-aligned therapeutic strategy is combination / multi-node intervention: upstream antigen depletion (R1) plus downstream amplifier modulation (H6 mitochondrial support, H7 autonomic stabilization, H8 mast-cell blockade) in protocolized combinations. The single counter-example to this generalization is the SIM01 → X.Dysbiosis intervention (Lau 2024 PMID 38071990; §7), which produces large composite-symptom effect sizes because X.Dysbiosis is upstream of multiple downstream amplifier pathways — it is not a single node, it is a root-level perturbation whose effects propagate through the same hierarchical structure CF-2 describes. SIM01 is not a refutation of CF-2; it is an exception that the architecture predicts.

Third, CF-2 has trial-design implications. The atlas predicts that conventional placebo-controlled monotherapy trials on single nodes will systematically under-power, because the mechanism’s effect on aggregate symptom burden is bounded by the ≤ 1.2 % single-node ceiling. Multi-arm factorial designs (e.g., a 2×2 antigen-depletion × autonomic-stabilization design) are the methodologically correct response. The HEAL-LC composite-mechanism endpoint trial (§6.4.2, §12) is the closest existing approximation; if it reads out with a positive composite endpoint but null per-component endpoints, that pattern would be the trial-readout signature of CF-2.

Fourth, CF-2 has methodological implications for the manuscript’s own reasoning. The four-criterion convergence is not independent: T1 and T8 partially share input data (the link-card primary citations), and T3 and T4 partially share computational machinery (both rely on the polytope topology). The honest statement is that CF-2 is a convergence across four partially-dependent analyses, not a four-fold independent replication. We discuss this caveat in §10 (Limitations) and frame CF-2 as a robust structural inference rather than a statistical replication. Reviewers should expect the CF-2 framing to be challenged on this dependence grounds; the response is that the structural conclusion (no single node has a manipulable knockout effect > 1.2 %) is invariant under reasonable reweighting of the four input techniques (see §11.5 anticipated reviewer challenges, CN-7).

§5.9 Disagreements between techniques (T11 D-1 through D-9)

The Phase 6 T11 capstone synthesis (disagreement-table.md) catalogues nine technique-vs-technique disagreements detected across the eleven analytical workstreams. Six dissolve under closer inspection as “different framings of the same question” rather than “different answers to the same question”; three are substantive disagreements requiring arbitration:

• D-1 T1 (BN noisy-OR analytic) vs T12 (Phase 5b directed pairs) on H7 → H8 and H7 → H12 cycles. Arbitrated: TRUST T12. The T1 author’s own audit memo §5.3 acknowledges that the DAG-acyclicity assumption fails for feedback loops; T12 verdicts each direction independently with PMID-anchored mechanistic evidence. The atlas annotates H7 ↔︎ H12 as a true cycle and the four asymmetric near-cycles (§4.5). Resolved.
• D-2 First T2 MR framing (ABO → LC P < 10⁻¹¹) vs empirical T2 run (β = +0.27, p = 0.027). Arbitrated: TRUST empirical run. The earlier P < 10⁻¹¹ figure does not appear in the audit memo and is not reproducible from the Lammi 2025 supplementary tables; the empirical β / p is reproducible with seed = 42. Documented as correction in the synthesis audit memo. Resolved (D-2 correction).
• D-3 T1 MCMC composite ranking vs T1 analytic knockout-impact ranking. Arbitrated: BOTH correct for their respective questions. MCMC composite answers “which mediators concentrate the most causal information flow”; analytic knockout-impact answers “which mediators if surgically removed most damage the strongest-path probability.” Different metrics, not a real disagreement. Both rankings are reported (§4.2, §5.1).
• D-4 T8 (mechanism continuity with ME/CFS) vs T7 (LC-specific RCT asymmetry). No disagreement. T8 addresses mechanism continuity at 12 mo (downstream); T7 addresses initiation-window specificity (upstream). Compatible: COVID-specific acute pharmacological-prevention signal upstream + final-common-pathway continuity downstream.
• D-5 T9 do-calculus identifiability vs T1 BN counterfactual leverage on H3.L10. Arbitrated: TRUST T9 for clinical recommendation; T1 for hypothesis weight. H3.L10’s +27.3 pp BN leverage is real but does not automatically translate into a causally-defensible therapeutic recommendation while the front-door assumption is contested. Honest framing: hold the clinical claim until HEAL-LC NCT07597902 reads out (PR-1; estimated 2027).
• D-6 T5 vulnerability score H6 = 8.5 vs T10 NC verdict “H6 is the only chain that PASSES NC.” No disagreement. Two methods converge on H6 as the methodologically strongest chain.
• D-7 T1 MCMC PPC r = 0.984 vs T1 author audit caveat. No disagreement — honest disclosure. The author has flagged that Beta-Binomial conjugacy yields automatic PPC accuracy; the synthesis adopts the framing and warns readers not to over-interpret PPC as external validation.
• D-8 H12 reclassification threshold not formally crossed by Y2 alone. Recommend reclassification as [CONSISTENT] not [ESTABLISHED]. Per honest-gap policy; three corroborating techniques (T12, T1 BN inter-root posteriors, T8 Sheehan analog) support architectural move while one PROBABLE_CAUSE incoming is short. Phase 7 disposition PR-2; v3.3 corpus refresh + Phase 5b rerun on H12 incoming is the trigger event for upgrade to [ESTABLISHED].
• D-9 T8 H11 verdict (“meta-modeling claim”) vs T10 H11 verdict (“structural NC paucity”). No disagreement — same finding, two framings. T5 manuscript-defense talking points reconcile via H11-as-paper vs H11-as-biological-claim distinction (§4.7, §2.4).

The disagreement-detection discipline is itself a methodological contribution: distinguishing genuine technique-disagreements from framing-disagreements is a learnable skill, and the v3.3 atlas should pre-register the exact substantive claim each technique aims to test (§2.7).

---

§6 Therapeutic identification

§6.1 The NMA topology — star around placebo, zero closed loops (T7)

The T7 network meta-analysis layer (analysis/network-meta/intervention-catalog.json) catalogues 42 candidate interventions and tests the LC trial corpus for network-MA feasibility. The catalogue covers antivirals (Paxlovid, Tenofovir, BC007), immunomodulators (rituximab, IVIG, anti-IL-6R, anti-IFN-α, IMC-1 valacyclovir + celecoxib, low-dose naltrexone, baricitinib), microbiome interventions (SIM01, FMT, butyrate, prebiotic blends), endothelial/coagulation interventions (apheresis HELP, anticoagulants, BC007 aptamer), autonomic / mast-cell interventions (cromolyn + H1 + H2 + montelukast factorial, ivabradine, fludrocortisone), neuroinflammation interventions (HBOT, intranasal insulin), and rehabilitation (pacing, graded-exercise, neurostimulation, cognitive remediation).

The empirical signature of the LC trial corpus is rendered as exhibit G33: a star graph centered on placebo, with 9 placebo-bridged edges, 0 head-to-head intervention-vs-intervention edges, and 0 closed loops. Without closed loops, no consistency test is possible, and no SUCRA / P-score can be defensibly computed (CF-4). The atlas explicitly refuses to publish a SUCRA ranking; this is the right epistemic move under the field’s current trial-corpus topology.

The 9 placebo-bridged RCT edges are dominated by R1-acute (Paxlovid acute prevention: HR ≈ 0.74 across three replications — Xie/Al-Aly 2023 JAMA Internal Medicine PMID 36951829, Bramante 2023 Lancet ID PMID 37302406, Ioannou 2023 Ann Intern Med PMID 37903369), R1-LC (Paxlovid established-LC: 4 sequential nulls — STOP-PASC PMID 38848477, PAX-LC, RECOVER-VITAL, Yale RCT), R2 (SIM01 synbiotic: Lau 2024 Lancet Infect Dis positive PMID 38071990), H10 (HBOT case-series-positive, RCT mixed), and H8 (cromolyn + montelukast small RCT mixed). Per-class effect estimates are rendered as exhibit G34 with the explicit caveat that no across-class pooling is defensible.

The 42-intervention catalogue, by mechanism class. The catalogue (analysis/network-meta/intervention-catalog.json) decomposes into seven classes with the following one-line evidence summaries:

• Antivirals (n=5) — Paxlovid is the dominant member; acute-prevention frame carries the field’s most-replicated positive RCT signal (HR ≈ 0.74 × 3), and the established-LC treatment frame carries the field’s most-replicated null RCT signal (4 sequential nulls). The hierarchical-DAG model resolves the apparent paradox: the upstream initiation node is interceptable; the downstream established-LC maintenance loop is not interceptable by monotherapy at the upstream node. Tenofovir + emtricitabine and BC007 (GPCR-autoantibody-targeting aptamer) round out the class as identifiable-but-empirically-thin candidates.
• Immunomodulators (n=9) — IVIG (NCT05445830 IVIG-SFN signal-positive in autoantibody-positive SFN subset; PMID 38311655), rituximab (no completed LC RCT), anti-IL-6R (RECOVER-IL6 pending), anti-IFN-α (no LC RCT), IMC-1 (FORTRESS null in n=422 LC same-sponsor replication; HEAL-LC NCT07597902 pending), low-dose naltrexone (mixed observational), baricitinib (no LC RCT), anti-CD20 (no LC RCT), tocilizumab (acute COVID only). The class is identifiable in principle (T9 Pearl) but RCT-deplete in LC-specific form.
• Microbiome interventions (n=5) — SIM01 synbiotic (Lau 2024 positive primary endpoint, PMID 38071990 — the keystone), FMT (no LC RCT; ME/CFS observational supportive), butyrate supplementation (no LC RCT; mechanism-based), prebiotic blends (no LC RCT), and Akkermansia supplementation (early data). SIM01 is the single class member to pass all five conditions; the others are mechanism-supported but Pearl-unidentified or RCT-thin.
• Endothelial / coagulation interventions (n=6) — HELP-apheresis (uncontrolled case-series only; the most-requested LC therapeutic by patient advocacy; §12 sham-RCT priority), anticoagulants (no completed LC RCT), BC007 aptamer (German uncontrolled cohort), nattokinase (early data), pentoxifylline (observational), and ozone therapy (no peer-reviewed evidence; cited only for completeness of the catalogue). The class is mechanistically rich (the microclot biology of Pretorius/Kell PMIDs 33203441, 39723824, 39942772, 39434957) but RCT-deplete.
• Autonomic / mast-cell interventions (n=7) — Cromolyn + H1 + H2 + montelukast factorial (small RCT mixed; §12 powered factorial priority), ivabradine (POTS data extrapolated), fludrocortisone (aldosterone-stratified subset hypothesis, §12 priority), midodrine (uncontrolled), propranolol (POTS data), pyridostigmine (off-label), and intranasal hydroxocobalamin (one-arm only). The class is the most clinically-prescribed at the bedside but lacks the placebo-controlled trial topology to support across-class ranking.
• Neuroinflammation interventions (n=5) — HBOT (RCT mixed at LC; positive in 5-session series, null in 40-session series), intranasal insulin (small RCT), low-dose lithium (mechanistic but no LC RCT), N-acetylcysteine (no LC RCT), and luteolin (preclinical only). The class lacks Pearl-identifiable LC-specific RCT readouts.
• Rehabilitation interventions (n=5) — Pacing / activity management (the standard of care; observational supportive), graded-exercise therapy (controversial; PEM-aggravating in subsets), cognitive remediation (small RCT positive), vagus-nerve stimulation (early data), and pulmonary rehabilitation (positive in cardiopulmonary-dominant LC subsets). Identifiability is generally good (interventions are randomisable); cross-domain validation is supportive but the LC-specific signal is heterogeneous across phenotypes.

Across the 42, only one — SIM01 — simultaneously satisfies the full Pearl + RCT + cross-domain stack (§7).

Five chains have zero LC-specific placebo-controlled RCT evidence (CF-8): H5, H7, H8, H12, H13. These five chains are biologically central but trial-empty in LC-specific form. The asymmetry is the central trial-design liability of the field; closing it is the dominant v3.3 trial-priority signal (§12.2).

§6.2 The Pearl counterfactual simulation (T9)

The T9 Pearl do-calculus counterfactual simulation (synthesis §7 narrative; the canonical T9 record is the synthesis §7 paragraph; T9 was not emitted as a standalone JSON) analyzed ten highest-leverage causal targets: 5 SPOFs from T4 (H3.L10, H10.L18, H11.L13, H11.L16, plus the pre-Phase-6 H11.L20 entry) and 5 mediator hubs from T1 (X.Dysbiosis, X.Histamine, X.Tcell_exhaustion, X.SCFA, X.Complement). For each target, the analysis tested back-door criterion satisfiability, front-door criterion satisfiability, latent-confounder enumeration, and intervention-priority score P = E × I × V (effect-weighted by counterfactual posterior shift, identifiability by Pearl criteria, validation by completed RCT + cross-domain).

Ten-target intervention-priority ranking (exhibit G31):

Rank	Intervention target	P score	Identifiability	RCT readout	Cross-domain
1	SIM01 → X.Dysbiosis	9.00	back-door ✓ + front-door ✓	positive (Lau 2024)	VALIDATED
2	H1+H2 blockade → X.Histamine	1.96	back-door ✓	no RCT	VALIDATED
3	IVIG → H10.L18 / H7	1.82	back-door ✓ (in trial)	case series only	PARTIAL
2b†	H3.L10 IMC-2 + Paxlovid (Pridgen combination antiviral)	3.75	front-door contested (D-5)	FORTRESS pooled null contested on two axes — protocol-specific (IMC-1 monotherapy ≠ IMC-2+Paxlovid combination) and pandemic-era cohort-1 confound, with FORTRESS cohort-2 (post-vaccine) meeting primary + secondary endpoints per Pridgen personal communication 2026-06-05; Pridgen 2017 FM RCT positive (PMID 28260944, n=143); Pridgen 2026 LC case series positive (PMID 41562079, n=24, d=1.8, 731-d durability); HEAL-LC pending	CONTESTED-POSITIVE
5	Butyrate / SCFA supplementation	1.12	front-door	observational only	VALIDATED
6	anti-C5 → X.Complement	0.50	back-door ✓	no LC RCT	INSUFFICIENT
7	T-cell exhaustion modulation	0	not identifiable	no LC trial	INSUFFICIENT
8	anti-spike Ab passive transfer	0	not identifiable	no LC trial	INSUFFICIENT
9	microclot apheresis (HELP)	0	back-door ✓ (in trial)	uncontrolled case series	LC_SPECIFIC
10	fludrocortisone (LC-POTS aldo-low)	0	back-door ✓ (in trial)	no LC RCT	PARTIAL

The priority ranking is dominated by SIM01 (P = 9.00, 2.4× the new runner-up). After the Pridgen-evidence ingestion (§6.5, §10.X) and the Phase 11e cohort-disclosure addendum (Pridgen personal communication, 2026-06-05), the H3.L10 row is re-scored at P ≈ 3.75 as the IMC-2 + Paxlovid combination protocol rather than as IMC-1 monotherapy, reflecting: (a) the positive Pridgen 2017 FM RCT (PMID 28260944, N=143, double-blind, multi-site, NCT01850420), an adjacent infection-associated chronic-illness indication; (b) the Pridgen 2026 LC-specific case series (PMID 41562079, Frontiers in Immunology, Jan 2026, n=24, IP vs IO p<0.0001, Cohen’s d=1.8, durability confirmed to 731 days under IRB-approved prospective consent); (c) the independent COX-2 antiviral mechanism with its own celecoxib-COVID RCT (Ghaznavi 2022 PMID 35834150 Inflammopharmacology; PGE2-storm mechanism Hong 2020 PMID 33312125) and the HSV/EBV-reactivation literature (Liu 2014, Higaki 2009, Gebhardt 2005, Gandhi 2015); the FORTRESS null in n=422 LC remains contested because (i) it tested IMC-1 monotherapy, not the IMC-2+Paxlovid combination protocol currently advocated, and (ii) HEAL-LC NCT07597902 is the pending definitive RCT test. IMC-2+Paxlovid moves from prior #4 to #2 in the priority ordering, above anakinra (P=1.96). The Phase 11e Pridgen personal-communication disclosure (2026-06-05) further refines the FORTRESS interpretation — cohort 2 (post-vaccine rollout) met both primary and secondary endpoints while cohort 1 (pandemic-era, vaccine-naive) drove the pooled null — motivating the modest V bump from 0.73 → 0.78 and the composite P move from 3.50 to 3.75 (§6.4.2 subsection d); the row is inserted in the table above (†2b) without renumbering the remaining rows because the v3.2.1 strict-canon T9 ranks #3–10 are referenced elsewhere in the manuscript. The remaining seven below-runner-up targets are either identifiable-but-empirically-thin (H1+H2 blockade, IVIG-SFN, butyrate), unidentifiable (T-cell exhaustion modulation, anti-spike passive transfer), or contested at the LC-RCT level. The atlas’s central honest therapeutic claim remains that one intervention — SIM01 — earns the full Pearl + RCT + cross-domain pedigree at the v3.2.1 strict canon, with IMC-2+Paxlovid as the next-best-evidenced candidate clearing four of five Pearl-rubric conditions pending the HEAL-LC readout (§6.5, §13).

Counterfactual methodology — how T1 posterior + cycle iteration produced the 9-intervention ranking. The T9 simulation drew its candidate target set from two sources: the four post-Phase-6 T4 SPOFs (H3.L10, H10.L18, H11.L13, H11.L16) plus the pre-Phase-6 H11.L20 entry retained for continuity, and the top-5 mediator hubs in the T1 MCMC composite ranking (X.Dysbiosis, X.SCFA, X.Histamine, X.Tcell_exhaustion, X.Complement). For each candidate, the simulation executed (i) a back-door identification pass against the 482-edge canonical DAG, enumerating the minimal sufficient adjustment set or returning UNIDENTIFIABLE; (ii) a front-door identification pass on the candidate intervention → mediator → outcome path; (iii) a latent-confounder enumeration listing the six recurring U-variables; (iv) an effect-weighted intervention-priority computation P = E × I × V, where E is the counterfactual posterior shift in symptom probability if do(X) is applied (drawn from the T1 MCMC counterfactual posterior with cycle-broken edges acknowledged), I is the identifiability tier (1.0 back-door + front-door; 0.7 back-door only; 0.4 front-door only; 0 unidentifiable), and V is the validation tier (1.0 positive RCT + cross-domain; 0.7 RCT only or cross-domain only; 0.4 observational only; 0 no evidence).

The cycle-iteration step is structurally important. The H7 ↔︎ H12 cycle and the four asymmetric near-cycles (§4.5) violate the T1 DAG-only assumption. The T9 simulation handles this by running a second-pass counterfactual on cycle-broken edges with the broken edges’ contributions held at posterior median and the rest of the network re-sampled; the difference between first-pass and second-pass posterior shifts is reported as a sensitivity bound on the effect-weight estimate. For SIM01, the cycle-iteration sensitivity is negligible (X.Dysbiosis sits upstream of the H7/H8/H10/H12 central-amplifier cluster and is not affected by the cycle edges). For X.Histamine and the H10-mediator pathways, the cycle-iteration sensitivity is non-negligible and is reported in the per-target sensitivity column of analysis/counterfactual/per-target-do-calculus.json.

The priority calculation is conservative: any candidate scoring P > 1.0 is treated as priority-grade; the SIM01 P = 9.00 result reflects the 4.6× gap to the runner-up. The atlas does not publish a SUCRA-equivalent rank across the 42 catalogued interventions; the T9 ranking is reported only for the 10 highest-leverage candidates where the methodology is defensible.

Six recurring latent confounders block observational identifiability of multiple targets and are the v3.3 cohort-measurement priority (§12.3):

• U_HLA — blocks X.Tcell_exhaustion and X.Complement identifiability.
• U_diet_lifestyle — blocks X.Dysbiosis and X.SCFA (the SIM01 trial randomisation breaks this within-trial, which is the structural reason SIM01 P = 9.00 while the next-best candidate sits at P = 1.96).
• U_sex_hormones — blocks X.Histamine; modulates H3.L10.
• U_HSV_susceptibility — blocks H3.L10.
• U_individual_vagal_tone — blocks H11.L13.
• U_baseline_microvascular_health — blocks H11.L16.

These six are the cohort-measurement priority for any future prospective LC cohort that aims to deliver Pearl-defensible observational evidence at the SIM01 standard.

The per-symptom relief landscape is rendered as exhibit G32: SIM01 dominates the GI, fatigue, brain fog, and concentration symptom domains (where Lau 2024 reported positive primary endpoint), with secondary contributions from X.SCFA (downstream of X.Dysbiosis) and X.Histamine (parallel to the gut-MC-axis), and gaps in the autonomic / SFN / sleep domains where the intervention landscape is empirically thin.

§6.3 CF-5: SIM01 → X.Dysbiosis as the single Pearl-defensible therapeutic claim

The central convergent finding of v3.2 is that SIM01 synbiotic → X.Dysbiosis is the only intervention surviving the full Pearl + RCT + sensitivity + cross-domain + cross-method stack (CF-5). The result is independently supported by:

• T9 Pearl do-calculus identifiability: SIM01 has back-door criterion satisfied directly by randomisation in the Lau 2024 RCT (PMID 38071990); front-door criterion additionally satisfied because the trial measured downstream gut microbiome composition via 16S sequencing.
• T7 NMA placement: Lau 2024 is one of 5 placebo-controlled positive RCTs in the LC corpus; primary endpoint positive on multiple measures (fatigue, GI, brain fog, sleep, well-being); ROBINS-I low-moderate.
• T1 MCMC posterior: X.Dysbiosis is the rank-#1 mediator by composite influence (0.234) in the MCMC ranking.
• T3 Cinelli-Hazlett robustness: the R2 backbone scores HIGH_ROBUST under heuristic; the LPS-translocation-to-systemic-inflammation axis has 10 cards with multi-group replication.
• T8 cross-domain ME/CFS validation: Giloteaux 2016 PMID 28095889 and Xiong 2023 PMID 36450980 document the same X.Dysbiosis mediator pathway in ME/CFS; cross-domain VALIDATED.

Every one of these five independent stress tests passes simultaneously. No other v3.2 therapeutic candidate clears all five. SIM01 → X.Dysbiosis is the central honest therapeutic claim of the v3.2 atlas, and §7 develops the keystone in depth.

§6.4 CF-8: Five chains are intervention-deplete

Five chains have zero LC-specific placebo-controlled RCT evidence: H5 (autoimmunity), H7 (small-fiber neuropathy / vagal), H8 (mast-cell hyperactivation), H12 (HPA axis), H13 (pediatric LC). These chains are biologically central but trial-empty in LC-specific form. The asymmetry vs R1 (5 placebo-controlled RCTs spanning acute prevention and established-LC treatment) is the central trial-design liability of the field.

The pattern is not random. H5, H7, H8, H12, H13 are concentrated in the “central downstream amplifier” tier and the pediatric specialization — chains where the underlying biology is well-mechanistically supported by cohort and case-series evidence but where the trial-design challenge is structurally hard: H5 requires autoantibody-positive cohort stratification at enrolment (the BHC-202 program is the first to attempt this at scale); H7 requires small-fiber-neuropathy-positive cohort stratification at enrolment (the IVIG-SFN extension is the only completed RCT in the chain); H8 requires mast-cell-mediator-positive stratification; H12 requires aldosterone- or cortisol-stratified stratification; H13 requires pediatric cohort infrastructure that does not yet exist at LC scale. The intervention-depletion is a downstream symptom of cohort-stratification infrastructure deficit, not a downstream symptom of mechanism uncertainty. The v3.3 priority queue (§12) targets each gap with the specific stratified-RCT design that the chain’s biology requires.

The asymmetry is also a peer-review-defensibility hazard. A reviewer at a top-tier journal can challenge any therapeutic claim resting on a chain in the intervention-deplete set with the entirely fair question, “where is the placebo-controlled RCT?”. The honest answer in v3.2 is that the trial does not exist; the chain rests on cohort + mechanism + cross-domain evidence; the v3.3 priority queue commits the field to closing the gap. This is the right posture, and the manuscript foregrounds it rather than hedging.

The v3.3 trial-priority list (§12.2) names the specific RCTs that would close each gap:

1. Sham-controlled HELP-apheresis RCT for R3 (the most informative single trial; the ThT-microclot-as-causal-node falsification test).
2. Powered H1+H2 LC-MCAS factorial trial (cromolyn + H1 + H2 + montelukast + LDN in MC-biomarker-positive LC-fatigue/PEM patients; T9 X.Histamine priority).
3. LC-SFN IVIG extension RCT (extending NCT05445830; H10.L18 SPOF + H7 chain coverage).
4. Iwasaki passive-transfer independent replication + autoantibody-targeted RCT (H5 worst-RoB-load chain).
5. Powered fludrocortisone-vs-placebo in aldosterone-stratified LC-POTS (H12.L08 falsification trigger unfulfilled).
6. Prospective pediatric growth/pubertal LC cohort + pediatric RoB pass (H13 98 unscored PMIDs).
7. HEAL-LC NCT07597902 readout watch (D-5 disagreement arbitration on H3.L10).

The intervention-priority list is constrained by both the T9 Pearl identifiability layer (only interventions with at least one identifiable Pearl criterion qualify for trial priority) and the T7 NMA trial-design topology (the field’s RCT capacity is currently consumed by R1-acute, leaving the other chains under-resourced).

§6.4.2 Pending trial readouts — what each will inform

Five pending or imminently-readable LC trial programs will most directly inform the v3.3 atlas refresh. Each is summarized below at ≈ 100 words, identifying which atlas chain or claim the readout most directly informs and what the readout-conditional Bayesian update would be.

FORTRESS / IMC-1 valacyclovir + celecoxib monotherapy and the broader Pridgen combination-antiviral protocol (H3.L10). The FORTRESS Phase IIb trial of IMC-1 in n = 422 LC patients reported a null primary endpoint in 2024, replicating the earlier same-sponsor null. The atlas’s H3.L10 card is in PR-1 disposition: T9 finds the front-door identifiability contested (D-5), and the FORTRESS null is the dominant negative evidence at the IMC-1 monotherapy card level. The broader H3 evidence package, however, is not coextensive with the FORTRESS readout. Three additional bodies of evidence inform the H3 chain after the v3.2 Pridgen-evidence ingestion (§6.2 priority-queue update; §10.X case-series-exclusion limitation):

(a) Pridgen 2026 LC case series. An IRB-approved prospective open-label case series of IMC-2 (valacyclovir 1500 mg BID + celecoxib 200 mg BID for 120 days) ± a 15-day Paxlovid course in n = 24 chronic-tail LC patients reported a 2-point IP-vs-IO advantage on PGIC fatigue at 120 days (p < 0.0001, Cohen’s d = 1.8), with comparable effect sizes on dysautonomia and brain-fog secondary endpoints, and durability confirmed across 120-, 305-, and 731-day post-protocol follow-up windows without statistically significant decay (Pridgen et al., Front Immunol 2026, PMID 41562079). The series carries the usual case-series caveats (no randomisation, no placebo, no blinding) and the usual case-series strengths (prospective consent, pre-specified primary endpoint, validated PGIC instrument, multi-timepoint durability, treatment-effect direction unambiguous by any standard). Time-from-infection to treatment averaged ~ 1 year (chronic-tail LC, exactly the population where Paxlovid monotherapy has failed in target-trial-emulation studies).

(b) Independent COX-2 antiviral mechanism. Celecoxib at the FDA-OA-approved 200 mg BID dose exerts a documented antiviral effect independent of its anti-inflammatory action, via two routes: (i) COX-2 inhibition disrupts the HSV-1 / EBV reactivation cycle through PGE2 suppression and EP-receptor signalling modulation (Liu et al. 2014 PMID 24688447 ScientificWorldJournal; Higaki et al. 2009 PMID 19274523 Curr Eye Res; Gebhardt et al. 2005 PMID 15857277 J Ocul Pharmacol Ther; Gandhi et al. 2015 PMID 26057147 Virology); and (ii) SARS-CoV-2 PGE2-storm reversal, where SARS-CoV-2 infection upregulates COX-2 and downregulates PGE2-degrading enzymes (Hong et al. 2020 PMID 33312125 Front Pharmacol), and the Ghaznavi 2022 COVID-19 RCT (PMID 35834150 Inflammopharmacology) reported celecoxib-arm mortality reduction in moderate-severe COVID-19. The combination of these two routes gives the IMC-2 protocol a second, independent antiviral mode of action layered on top of the valacyclovir herpesvirus suppression and the Paxlovid SARS-CoV-2 Mᵖʳᵒ inhibition (Deng et al. 2014 PMID 25100843 J Virol documents the coronavirus 3CLpro inhibitor fitness landscape that underpins the Paxlovid arm). Independent corroboration of the sustained-antiviral protocol rationale is provided by Peluso et al. 2023 (medRxiv preprint, doi:10.1101/2023.07.27.23293177), which used multimodal molecular imaging to demonstrate tissue-based T-cell activation and SARS-CoV-2 RNA persistence up to 2 years post-infection — the exact mechanism the 120-day antiviral protocol is designed to target.

(c) Protocol-specific vs mechanism-specific interpretation of FORTRESS. The FORTRESS null is therefore best read as a protocol-specific finding — IMC-1 monotherapy in chronic-tail LC under the FORTRESS dose/duration/cohort — rather than as a falsification of the broader composite-antigen-amplification mechanism story. The combination IMC-2 + Paxlovid protocol that Pridgen 2026 advocates is empirically distinct from FORTRESS at the dose (celecoxib 200 mg BID vs FORTRESS dose), duration (120 days vs 14 days), and regimen (combination antiviral with Paxlovid pulse vs monotherapy) levels.

(d) Pandemic-era cohort-1 vs cohort-2 stratification within FORTRESS. In a 2026-06-05 reply email to the atlas team, Dr. W.L. Pridgen disclosed a second, internal-to-FORTRESS axis of contestation against the pooled null. Per Pridgen, the pandemic substantially affected the FORTRESS trial at the cohort level: the first cohort failed to meet statistical significance, but the second cohort, enrolled after the vaccine was introduced, was less affected by pandemic-era confounding and met statistical significance in both the primary and the secondary endpoints. Pridgen’s own retrospective assessment was that the study “should have stopped until the pandemic was through.” The implied confounders for cohort 1 are the high incident SARS-CoV-2 re-infection rate during the trial enrollment window, vaccine-naive participants, and variant churn contaminating the placebo arm; cohort 2 was enrolled under cleaner background-infection conditions and met its endpoints. Provenance: W.L. Pridgen, personal communication, 2026-06-05 (reply email). The v3.3 atlas commits to cross-checking this disclosure against any published cohort-stratified FORTRESS reanalysis when one becomes available. The corresponding scoring adjustment is rubric-disciplined: the V-score for H3.L10 is bumped modestly from 0.73 → 0.78 (composite P from 3.50 → 3.75, rank #2 unchanged), crediting the cohort-2 positive signal without earning full Pearl-rubric credit until a published cohort-2-only analysis lands. This is itself a worked example of the case-series-tier-vocabulary question raised in §10.12.2: a personal-communication trial-cohort-disaggregation disclosure occupies a tier that v3.2’s evidence-vocabulary does not yet name, and v3.3’s expanded case-series-tier vocabulary should cover it explicitly.

HEAL-LC NCT07597902 readout (pending 2026) remains the definitive RCT-grade test of the combination protocol in LC. A positive HEAL-LC readout would re-promote H3.L10 to CONSISTENT and would place IMC-2 + Paxlovid alongside SIM01 at the top tier of Pearl-defensible LC therapeutics; a null readout would demote H3.L10 to SPECULATION at the strict-canon level and prompt a Pridgen-2026 effect-size sensitivity audit. Until HEAL-LC reads out, the atlas treats IMC-2 + Paxlovid as the second-best-evidenced therapeutic candidate in the v3.2 corpus (§6.2, P ≈ 3.75 after the Phase 11e cohort-disclosure addendum).

HEAL-LC NCT07597902. The composite-mechanism endpoint design (combined effect on viral persistence biomarkers + symptom relief + neurocognitive performance) is structurally novel for LC trials. Positive readout would validate the hierarchical-amplification mechanism of H3 (composite antigen / IFN-I drive) and would be the first LC RCT to demonstrate an identifiable Pearl-defensible effect outside SIM01. The atlas’s posterior shift if HEAL-LC reads positive: H3.L10 promotes to CONSISTENT or ESTABLISHED; the four-chronic-gates field-diagnosis update credits c4 (CI excluding null) for the chain; the v3.3 priority queue de-escalates the chain.

IVIG-SFN extension (NCT05445830). Vernino 2024 PMID 38311655 reported the first IVIG-SFN trial in autoantibody-positive small-fiber-neuropathy LC patients with mixed signal: the autoantibody-positive subset showed clinical-meaningful improvement; the autoantibody-negative did not. The extension trial powers the autoantibody-positive subset analysis at n adequate for primary-endpoint inference. A positive extension would validate H7 + H5 combined mechanism in a stratified subset and would promote H10.L18 (the IVIG-target SPOF card) from EMERGING to CONSISTENT. The readout is the first definitive test of the autoantibody-stratified IVIG hypothesis in LC.

Sham-controlled HELP-apheresis RCT (proposed, §12). The most-requested LC therapeutic at the patient-advocacy level; the most-studied LC intervention at the uncontrolled-case-series level; the trial design that would convert the microclot hypothesis from “mechanism-supported” to “causally-testable” status. The trial does not yet exist in completed form; the proposed protocol (apheresis-vs-sham in n = 200 microclot-positive LC patients with PASC-12 primary endpoint at 6 months) is in v3.3 priority. A positive readout would convert R3 from biologically-supported / RCT-empty to RCT-anchored and would promote one or two R3 cards to ESTABLISHED. A null readout would demote the R3 microclot-as-causal-node claim from CONSISTENT to EMERGING.

BHC-202 GPCR-autoantibody-targeting program. The first LC trial targeting GPCR autoantibodies (β1-AR, M2-AChR, AT1R) in stratified autoantibody-positive cohorts. The program design is the cleanest current implementation of the H5-stratified-RCT principle and is the most likely near-term source of an H5 ESTABLISHED-tier promotion. Positive readout would validate the autoantibody-positive cohort-stratification approach to LC trials and would be a structural advance for the field’s trial-design infrastructure.

Readouts from FORTRESS / HEAL-LC / IVIG-SFN extension are expected in the 2026–2027 window and will trigger a v3.3 corpus refresh; the apheresis sham-RCT and BHC-202 readouts are the principal v3.4 anchors.

§6.5 The gap between mechanism strength and causally-defensible therapeutic recommendation

The atlas’s central honest therapeutic finding is that the gap between mechanism strength and causally-defensible therapeutic recommendation is the largest single gap in the v3.2 evidence layer. H3.L10 has +27.3 pp BN counterfactual leverage but front-door identifiability is contested (D-5, PR-1). X.Tcell_exhaustion ranks #4 in MCMC composite influence but is not Pearl-identifiable (no clean intervention exists). X.Complement ranks #5 in MCMC composite influence but the anti-C5 intervention is not Pearl-identifiable in LC-specific form (no completed LC RCT).

This gap is the empirical signature behind the central honest framing of v3.2’s therapeutic landscape: one claim survives the full bar; ten others fail at one or more conditions. The right epistemic posture is to commit fully to the SIM01 claim, to flag the other targets as identifiable-but-empirically-thin (where the structural fix is RCT funding) or unidentifiable (where the structural fix is cohort-measurement of latent confounders), and to refuse to publish a quantitative across-class intervention ranking that the field’s trial-corpus topology does not support.

Phase 11e cohort-disclosure refinement of the H3.L10 narrowing. The Phase 11d Pridgen-evidence ingestion narrowed the SIM01-vs-runner-up gap from 7× to ~2.6×; the Phase 11e Pridgen personal-communication addendum (2026-06-05) narrows it further to ~2.4×, with the H3.L10 IMC-2 + Paxlovid composite priority moving from P ≈ 3.50 to P ≈ 3.75. The trigger is the FORTRESS cohort-1-vs-cohort-2 disclosure summarised in §6.4.2 subsection d: cohort 2 (post-vaccine) met both primary and secondary endpoints while cohort 1 (pandemic-era) drove the pooled null. The interpretation discipline matters as much as the score: the V bump is modest (0.73 → 0.78) precisely because personal-communication trial-cohort-disaggregation disclosures, however interpretively decisive, are not yet a published cohort-stratified reanalysis. The v3.3 atlas commits to a further V update (and rank/score audit) if and when an independent published cohort-stratified FORTRESS reanalysis lands.

---

§7 SIM01 keystone

§7.1 The headline claim

Oral synbiotic SIM01 administration to PACS / Long COVID patients reduces the marginal probability of fatigue, brain fog, gastrointestinal symptoms, and concentration difficulty at six months versus placebo, mediated through restoration of gut microbiome composition and downstream SCFA production — and this is the only therapeutic claim in v3.2 that simultaneously survives Pearl’s identification criteria, completed placebo-controlled RCT evidence, mediator-hub influence ranking, Cinelli-Hazlett robustness bounds, and independent ME/CFS cross-domain validation.

This is the central honest therapeutic claim of the v3.2 atlas. The five conditions that SIM01 simultaneously meets are summarized in exhibit G35 (the SIM01-keystone five-conditions panel).

§7.2 Primary evidence — Lau 2024 Lancet Infectious Diseases

Citation. Lau RI, Su Q, Lau ISF, et al. A synbiotic preparation (SIM01) for post-acute COVID-19 syndrome in Hong Kong (RECOVERY): a randomised, double-blind, placebo-controlled trial. Lancet Infect Dis. 2024;24(3):256–265. PMID 38071990.

Design. Randomised, double-blind, placebo-controlled trial, single-centre (Chinese University of Hong Kong, CUHK). Adult PACS patients meeting WHO-criteria post-acute COVID-19 syndrome at ≥ 4 weeks post-acute infection. Randomised 1:1 to SIM01 oral synbiotic (proprietary multi-strain Bifidobacterium preparation plus three prebiotic compounds) versus matching placebo for 6 months. Primary endpoint composite of symptom resolution across multiple LC domains assessed at 6 months.

Primary endpoint result. SIM01 produced statistically significant improvement versus placebo on multiple symptom domains at 6 months: fatigue, gastrointestinal symptoms, memory/concentration (“brain fog”), and general well-being. Adverse-event profile mild and similar between arms; no serious adverse events attributed to study intervention. Risk-of-bias assessment: low-to-moderate (standard double-blinded RCT design, prospectively registered, intention-to-treat analysis; the principal RoB caveats are single-centre design and the proprietary nature of SIM01 limiting generic-formulation replication).

Mediator measurement. The trial measured downstream gut microbiome composition (16S sequencing on stool samples pre- and post-intervention) and reported restoration of microbiome diversity in the SIM01 arm versus minimal change in the placebo arm. This provides the front-door mediator measurement that Pearl identification requires.

§7.3 X.Dysbiosis at composite influence rank #1

X.Dysbiosis is the #1 mediator in the MCMC composite-influence ranking with composite_influence_score = 0.234 (analysis/bayesian-net/sensitivity-ranking.json). The four follow-up nodes are X.SCFA (0.114), X.Histamine (0.061), X.Tcell_exhaustion (0.049), and X.Complement (0.048). X.Dysbiosis carries roughly 2× the composite influence of the next-ranked mediator (X.SCFA), which itself sits downstream of X.Dysbiosis in the H9/R2 chain. The combined X.Dysbiosis + X.SCFA “gut axis” influence (0.348) exceeds the cumulative composite influence of all other top-5 mediators combined.

SIM01 targets precisely the highest-influence mediator hub in the v3.2 Bayesian network. The intervention-priority calculation P = E × I × V yields 9.00 for SIM01 versus 1.96 for the runner-up X.Histamine intervention — a 4.6× priority gap.

The analytic-canonical Bayesian run (Beta-Binomial conjugate) ranks knockout-impact differently — X.Tcell_exhaustion at 0.0112 with X.Dysbiosis lower-ranked because the strongest-path-probability metric depends on whether the chain initiation node is included. This is disagreement D-3, arbitrated as “both correct for respective questions.” For mediator-hub influence (the question SIM01’s claim turns on), the MCMC composite is the appropriate metric.

§7.4 R2 backbone robustness (T3) and R2.L11 as the SIM01 anchor

The R2 / H9 dysbiosis-chain backbone scores well under Cinelli-Hazlett: the LPS-translocation-to-systemic-inflammation axis has ten cards with multi-group replication, and the closest formula-RV-computable card in the R2 cluster (R2.L08 SCFA-deficit → mitochondrial impairment) yields a HIGH_ROBUST classification under the heuristic RV.

Closed-form Cinelli-Hazlett RV is not directly available for the Lau 2024 RCT itself (it is a randomised intervention, not an observational claim — Cinelli-Hazlett applies to confounder-robustness of observational estimates; the RCT is identifiable by back-door criterion directly). The robustness story for SIM01 specifically rests on the convergent evidence stack (§4 + §6 + §7 + Supplementary Appendix B of this manuscript), not on a Cinelli-Hazlett bound on the trial itself.

The Phase 7 strict re-tier audit promoted R2.L11 (gut LPS / cytokines + dysbiosis-derived metabolites → vagal-afferent sensitisation → autonomic dysregulation) from EMERGING to ESTABLISHED at score 6/8. R2.L11 is the single R-chain ESTABLISHED-tier survivor in the v3.2.1 strict-criteria-compliant canon (Supplementary Appendix B (relocated §7.5)). The Lau 2024 RCT is the prospective interventional anchor; the c1 PI-independence is satisfied by the CUHK Lau group plus the independent ME/CFS gut-dysbiosis validation literature. R2.L11 is the cornerstone of CF-5 and the structural backbone of the keystone-therapeutic core in the v3.2.1 canon.

§7.5 ME/CFS cross-domain VALIDATED (T8)

Two key anchor citations independently validate the X.Dysbiosis mediator pathway in chronic post-viral syndromes:

• Giloteaux L, Goodrich JK, Walters WA, et al. Reduced diversity and altered composition of the gut microbiome in individuals with myalgic encephalomyelitis/chronic fatigue syndrome. Microbiome. 2016;4:30. PMID 28095889. Documents reduced gut microbial diversity in ME/CFS patients versus matched healthy controls, with LPS-translocation markers (sCD14, LBP) correlating with symptom severity.
• Xiong R, Gunter C, Fleming E, et al. Multi-omics of gut microbiome-host interactions in short- and long-term myalgic encephalomyelitis/chronic fatigue syndrome patients. Cell Host Microbe. 2023;31(2):273–287.e5. PMID 36450980. Documents multi-omics-confirmed reduced butyrate-producing bacteria, altered tryptophan and SCFA pathways, and host-microbiome interaction signatures distinguishing ME/CFS from controls.

T8’s cross-domain validation matrix classifies the LC X.Dysbiosis / R2 chain as VALIDATED in the ME/CFS comparator and PARTIAL in the PTLDS comparator. The post-viral / dysautonomia / chronic-fatigue syndromes converge on the gut-dysbiosis-axis mediator pathway across multiple disease contexts. This is independent corroborative evidence for the mediator pathway SIM01 targets — not LC-specific external replication of the SIM01 trial itself, but mechanism-level validation of the target.

§7.6 Pearl do-calculus identifiability (T9)

Back-door criterion. The Lau 2024 SIM01 trial is a randomised double-blind placebo-controlled RCT. Randomisation severs all back-door paths from intervention assignment to outcome by construction. This is the textbook Pearl back-door identification — randomised assignment of treatment makes the do(SIM01) interventional distribution identical to the observed conditional P(outcome | SIM01 = assigned) within the trial sample.

Front-door criterion. The trial additionally measured the downstream mediator (gut microbiome composition via 16S sequencing on stool). This provides a front-door identifiability path through the X.Dysbiosis → X.SCFA → outcome mediator chain. Even setting aside the back-door identification (which alone is sufficient), the front-door criterion is independently satisfied because the mediator was directly observed and the trial documents both the intervention → mediator step and the mediator → outcome step.

Recurring latent confounders relaxed by RCT design. Outside the RCT setting, X.Dysbiosis identifiability is blocked by U_diet_lifestyle (synthesis §12 confounder list). Within the RCT, randomisation breaks U_diet_lifestyle’s confounding effect on outcome — the unobserved-confounder bias that would dominate any observational cohort estimate is eliminated. This is precisely why the trial result has the identifiability properties that the observational X.Dysbiosis association does not, and is the structural reason SIM01 sits at intervention-priority P = 9.00 while the next-best candidate sits at P = 1.96.

Identifiability verdict. Identifiable by both back-door (primary) and front-door (secondary) criteria. No other candidate intervention in the v3.2 atlas achieves this dual-criterion identifiability simultaneously with a completed positive primary-endpoint RCT and independent cross-domain validation.

§7.7 The five conditions — table

#	Condition	SIM01 status	Reference
1	Identifiable under Pearl’s do-calculus (back-door or front-door criterion)	✓ Back-door (randomisation) + front-door (mediator measured)	§7.6
2	Completed placebo-controlled RCT with positive primary endpoint	✓ Lau 2024 PMID 38071990	§7.2
3	Cinelli-Hazlett robustness on related claims	✓ R2 backbone HIGH_ROBUST under heuristic; LPS-translocation axis multi-group replication	§7.4
4	Cross-domain validation in independent comparator literature	✓ ME/CFS gut-dysbiosis VALIDATED (PMIDs 28095889, 36450980)	§7.5
5	High mediator-hub influence in the canonical Bayesian network	✓ X.Dysbiosis composite-influence rank #1 (0.234) in MCMC	§7.3

No other v3.2 therapeutic candidate clears all five conditions. The four unidentifiable candidates (T-cell exhaustion modulation, anti-spike Ab passive transfer, microclot apheresis without sham control, anti-NMDAR immunotherapy without LC RCT) fail condition #1. The three identifiable-but-empirically-thin candidates (H1+H2 blockade, IVIG-SFN, butyrate supplementation) fail condition #2 (no completed positive LC-specific RCT). The contested H3.L10 / IMC-1 / HEAL-LC candidate fails condition #2 (FORTRESS null in n=422 LC same-sponsor replication; HEAL-LC NCT07597902 pending) and additionally fails condition #1 conditionally (D-5 disagreement — front-door contested).

§7.8 Honest caveats

The Pearl-defensible verdict does not mean SIM01 is a panacea. Five honest caveats are explicit and propagated to §11 limitations:

1. Single trial, single centre. Lau 2024 is one RCT at one centre (CUHK Hong Kong). The most important follow-on study is independent-centre replication of SIM01 (or comparable synbiotic) in different LC cohorts — European, North American, South American, sub-Saharan African — to test geographic generalizability. Until at least one independent-centre replication reads positive, the claim is “supported by one well-conducted RCT” not “established beyond plausible doubt.”

2. Primary endpoint versus per-domain effect magnitudes. The Lau 2024 trial reported positive primary-endpoint composite results, but the per-domain symptom-domain effect magnitudes vary. Primary endpoint positivity does not imply equal benefit on all nine v3.2 polytope symptom domains. The manuscript does not over-extrapolate the per-domain effect sizes beyond what the trial directly reports.

3. Generalizability across LC phenotypes. The Lau 2024 cohort was Hong Kong PACS patients enrolled in 2022–2023, predominantly Omicron-era infections. Whether the SIM01 effect generalizes to LC phenotypes dominated by other variants (original Wuhan, Alpha, Delta) or to LC subgroups with predominant non-gut mechanisms (e.g., R3 microclot-dominant, H5 autoantibody-dominant) is empirically open.

4. Mechanism specificity — synbiotic versus FMT versus prebiotic alone. SIM01 is a proprietary multi-strain Bifidobacterium + prebiotic preparation. The trial does not decompose the contribution of (a) the bacterial strains specifically, (b) the prebiotic compounds specifically, (c) the synergy between them, or (d) more aggressive interventions (faecal microbiota transplant) that might achieve larger effects. The mechanism-level claim (“restoring gut microbiome composition reduces LC symptoms”) is supported; the formulation-level claim (“SIM01 specifically is the optimal intervention”) is supported only insofar as SIM01 is the formulation tested.

5. U_diet_lifestyle outside the RCT. SIM01’s identifiability rests on randomisation breaking U_diet_lifestyle confounding within the trial sample. Real-world deployment in non-randomised clinical care reintroduces the U_diet_lifestyle bottleneck. Observational post-marketing-surveillance studies of SIM01 effects in clinical care will be confounded by the diet-lifestyle factors the RCT randomisation removed. Future LC cohorts must measure validated dietary intake to relax this observational bottleneck.

§7.9 The contrast — 41 catalogued therapeutics fail ≥ 1 condition

The Phase 5 NMA catalogued 42 LC interventions across antiviral, immunomodulatory, microbiome, endothelial/coagulation, autonomic/mast-cell, neuroinflammation, and rehabilitation classes. Exactly one — SIM01 → X.Dysbiosis — survives all five Pearl-defensibility conditions. The remaining 41 fall into four failure-mode classes:

• Failure of condition #1 (not Pearl-identifiable in LC-specific form): anti-spike Ab passive transfer (no clean intervention), T-cell exhaustion modulation (no manipulable target node with measured mediator), anti-NMDAR immunotherapy (no LC RCT design), microclot apheresis without sham control (back-door satisfied within trial but no sham-controlled trial exists), HBOT in non-pre-specified cohorts (heterogeneous protocols, no clean back-door identification).
• Failure of condition #2 (no completed positive LC-specific RCT): rituximab (small case series only), IVIG-SFN (case series only), anti-IFN-α (no LC RCT), baricitinib (no LC RCT), butyrate / SCFA supplementation (observational only), FMT (small case series only), low-dose naltrexone (no completed placebo-controlled LC RCT despite multiple ongoing trials), fludrocortisone in LC-POTS (no completed RCT), cromolyn + montelukast combination (small mixed RCT only).
• Failure at condition #2 with contested condition #1 (FORTRESS null + HEAL-LC pending): H3.L10 IMC-1 (valacyclovir + celecoxib + Paxlovid). FORTRESS NCT04748705 read out null in n = 422 LC same-sponsor replication of the Pridgen 2017 fibromyalgia positive (PMID 28260944). HEAL-LC NCT07597902 is the antigen-positive-enriched LC follow-on test (PR-1 disposition).
• Failure of condition #5 (low mediator-hub influence in BN): rehabilitation / pacing / graded-exercise (heterogeneous targets, no single mediator pathway); cognitive remediation (downstream symptom, not mediator); neurostimulation (no clean mediator pathway).

The atlas’s honest position is that the gap between mechanism strength and causally-defensible therapeutic recommendation is wide. SIM01 is on one side of the gap; the other 41 candidates are on the other. Closing the gap is the trial-priority work of v3.3 (§12.2).

§7.10 Pending trial readouts

Five trial readouts are expected to materially update the v3.2 atlas within 12–24 months:

• FORTRESS NCT04748705 — H3 / IMC-1 (valacyclovir + celecoxib) in LC. Read null (n = 422, primary endpoint missed). The same-sponsor replication of the Pridgen 2017 fibromyalgia positive (PMID 28260944) failed in LC, which weakens the H3.L10 single-card claim but does not falsify the broader H3 composite-mechanism chain (which has antigen-positive-enriched sub-population predictions tested next by HEAL-LC).
• HEAL-LC NCT07597902 — Mt Sinai antigen-positive-enriched LC IMC-1 + Paxlovid combination. Expected readout 2027. PR-1 disposition: if HEAL-LC reads null, H3.L10 is downgraded to EMERGING; the +27.3 pp BN leverage is re-interpreted as model artifact. If HEAL-LC reads positive in antigen-positive subgroup, H3.L10 is re-promoted to ESTABLISHED, and the H3 composite-mechanism prediction is empirically validated.
• IVIG-SFN extension RCT (extending NCT05445830) — H10.L18 SPOF and H7 chain coverage. Pending design.
• Apheresis sham-RCT for R3 microclot — the ThT-microclot-as-causal-node falsification test. Pending design.
• BHC-202 / SIM01 v2 — independent-centre SIM01 (or comparable synbiotic) RCT replication. Pending design. The most important single trial for converting SIM01 from “Pearl-defensible with one positive RCT” to “Pearl-defensible with multi-centre replication.”

These five readouts will reshape the v3.2 atlas. The dual-version v3.2 / v3.2.1 architecture is constructed so that the atlas can absorb each readout without disturbing the existing canonical record: a positive HEAL-LC reads as a v3.2.1 promotion of H3.L10 + a parallel v3.3 atlas mutation; a positive SIM01 replication reads as a v3.3 promotion of R2 backbone cards + an additional evidence_tier_pre_phase8 annotation block.

§7.11 Captain’s framing

“One trial that survives full scrutiny is more publishable than fifty that don’t.” This is the operational principle behind the v3.2 atlas’s therapeutic-identification discipline. The atlas refuses to publish a quantitative across-class intervention ranking that the current LC trial-corpus topology does not support (CF-4). It refuses to claim Pearl identifiability for interventions whose back-door is contested (D-5, H3.L10). It refuses to inflate the SIM01 result by claiming multi-centre external replication that does not yet exist (§7.8 caveat 1). It does claim — and stakes the manuscript on — that the SIM01 → X.Dysbiosis intervention is the single Pearl-defensible therapeutic claim that emerges from the eleven analytical techniques applied to the 5,101-paper corpus. This claim is the keystone of the v3.2 atlas’s clinical narrative.

---

§8 The four-universal-fail field diagnosis

The per-chain demotion signature is shown in exhibit G53 — a 13-chain × 8-criterion heatmap demonstrating that the four-universal-fail pattern (C4 effect-size CI 96 %, C3 prospective design 81 %, C7 ROBINS-I-passing 73 %, C6 PI-independence 59 %) is invariant across all thirteen chains. The right-margin chain-average bars are flat; the bottom-margin criterion-average bars show the four universal-fail peaks. The pattern is not chain-specific; it is field-wide. Exhibit G53: Per-chain × per-criterion demotion-rate heatmap across the 107 Phase-7 demoted cards. The four universal-fail criteria — C4 effect-size CI (96% fail-rate), C3 prospective design (81%), C7 ROBINS-I-passing (73%), C6 PI-independence (59%) — are framed in DANGER red and persist at field-level frequencies across all 13 chains. Right-margin per-chain demotion-rate bars are approximately flat: the pattern is invariant across chains and is field-wide structural, not atlas-specific or curation-artifact.

§8.1 The pattern of demotions, not the demotions themselves

The Phase 7 strict-criteria audit demoted 107 cards. The number is large; the interpretation matters more. This is not a finding that the v3.2 atlas is bad. This is a finding about the Long COVID research base as a whole. Across the 107 demotions, four criteria fail to be credited at field-level frequencies of 59–96 percent. Each gate has a specific structural cause; closing each is a concrete v3.3 research-agenda item.

The atlas’s reading is that the Long COVID literature is a young, fast-growing, multi-mechanism research base that has not yet adopted the methodological infrastructure that older mechanistic literatures (e.g., cardiovascular trials, oncology biomarkers) take for granted. The cards demoted by Phase 7 are not necessarily wrong; many of them are mechanism-plausible and conceptually coherent (Supplementary Appendix B (relocated §7.1)). They simply lack the machine-readable evidence forms that strict peer-review filtering requires.

§8.2 The four chronic gates in detail

Gate c4 — machine-readable effect-size confidence interval (96.3 % not credited). This is the highest-frequency gate failure. 78 of the 103 non-credits are UNVERIFIABLE rather than False — meaning the atlas card encodes quantitative_parameters in narrative form (“Δ ACTH 247 %, peak cortisol 4.2 ng/dL post-stim”) without machine-readable 95 % CIs. The atlas inherited this style from primary literature: most cited papers report point estimates and p-values without bounding the CI, or they bound it in figure legends rather than abstract text. The structural cause: Long COVID journals do not enforce a 95 %-CI-with-null-clause reporting standard. The fix at the curation layer is to add an effect_size_95ci required schema field; the fix at the field layer is journal-level reporting-standard adoption. The v3.3 priority IM-12 (“machine-readable schema for effect sizes”) addresses this at the curation layer; the field-level fix is editorial.

Gate c2 — prospective design among primary citations (81.3 % not credited). Of 107 demotions, 65 cards have zero prospective-design primary citations. The LC literature is dominantly cross-sectional case-control: LC patients versus convalescent controls at a single late-postacute time point. The c2 gate correctly flags this as inadequate for causal-mechanism claims. The structural cause is funding and design: prospective LC cohorts are expensive (enrollment during acute infection then 12+ months of follow-up) and the field has only a handful (RECOVER PASC, FIND, German COVIM cohort). The 4 sequential null Paxlovid-LC RCTs are the highest-quality prospective interventional design in the field, and they read out null — a result that hardens the case for hierarchical-DAG framing but does not satisfy c2 for upstream mechanism cards. The fix: NIH / Wellcome / EU prospective-cohort funding extended to neuroinflammation (H10), autoantibody (H5), mitochondrial (H6), HPA (H12), and pediatric (H13) subspecialty cohorts (§12).

Gate c3 — ROBINS-I composite ≤ moderate (72.9 % not credited; 78 of 78 False, no UNVERIFIABLE). The ROBINS-I composite scoring finds that 78 of the 107 demoted cards cite at least one primary paper with critical-tier RoB. The structural cause is confounding by indication and selection: LC cohorts are typically self-referred or symptom-screened, and convalescent controls are typically asymptomatic at recruitment — a self-selected difference that confounds any biomarker comparison. The c3 gate correctly observes that this is not a per-card flaw; it is a field-wide confound that GRADE-style reviewers will downgrade for at every claim. The fix: ROBINS-I-tier-enforced cohort recruitment standards for LC-vs-control comparison studies, with active recruitment of matched-deconditioning and matched-chronic-illness controls (the H6 / Appelman-Wüst CPET-PEM design is the template).

Gate c1 — PI-independence ≥ 2 distinct groups (58.9 % not credited; 51 of 63 UNVERIFIABLE). A majority of the c1 non-credits are UNVERIFIABLE rather than False — meaning the atlas card has too few citations to determine PI-independence (typically 2–4 PMIDs, where the rubric requires ≥ 2 distinct PI groups demonstrated). The structural cause is single-PI domination of subfields: the Pretorius / Kell group on microclots, the Iwasaki group on autoantibodies, the Putrino group on dysautonomia, the Naviaux group on CDR / dauer. These groups produce excellent work, but a card that anchors solely or primarily on one of them cannot satisfy c1. The fix: NIH and equivalent funders need to fund out-of-PI-group replications of the dominant findings. The v3.3 priority IM-15 (NIH/Wellcome out-of-PI-group replication budget) addresses this at the field level.

§8.2.2 Worked case study — H11.L13 demotion CONSISTENT → EMERGING

The four chronic gates are most legible in a worked case study of a specific demotion. We choose H11.L13 — the vagus-nerve-inflammation card linking acute SARS-CoV-2 neurotropism to reduced cholinergic anti-inflammatory tone and downstream skeletal-muscle / IGF-1-axis dysregulation — because it sits at the intersection of three central-amplifier chains (H7, H10, H11) and because its biology is mechanistically rich but its evidence form fails four of the eight strict-rubric gates.

The mechanism. Acute SARS-CoV-2 infection invades the vagus nerve through olfactory-bulb / glossopharyngeal cross-talk and direct neural ACE2 expression. Monocyte infiltration of the vagal trunk drives a neuronal, endothelial, and Schwann-cell inflammatory transcriptional program. The downstream consequence is reduced cholinergic anti-inflammatory reflex tone (a well-characterized neuroimmune negative-feedback loop) and disruption of vagal trophic input to skeletal muscle, including IGF-1-axis dysregulation that may contribute to the post-exertional malaise phenotype. The mechanism is plausible. The card carries 6 primary citations across multiple PI groups (c1 credited).

Gate-by-gate audit. c1 PI-independence: credited (6 citations span multiple groups). c2 prospective design: not credited (only 1 of the 6 supporting citations has a documented prospective design tag in the canonical bibliography). c3 ROBINS-I composite ≤ moderate: not credited (only 0 of 1 RoB-scored supporting citation is moderate-or-lower). c4 effect-size 95% CI: not credited (sparse quantitative data; only 1 numeric effect-size estimate across the 6 citations, and that estimate is reported without 95% CI bounds). c5 cross-domain coherence: not credited (“LC_specific_no_external_validation” — the vagus-nerve-inflammation finding has no parallel in the ME/CFS, POTS, PTLDS, post-Ebola, or autoimmune comparator literatures at sufficient resolution to validate). c6 negative control: UNVERIFIABLE (“T10 = structural NC paucity”; no canonical NC data). c7 falsifier named: credited (the card encodes a concrete trial-or-experiment falsifier with null-clause). c8 sign consistency: credited (forward direction supported across 6 citations).

Score and verdict. Sum = 4/8 credited, with 1 UNVERIFIABLE that does not credit per the conservative scoring rule — effective score 3/8. The v3.2 tier was CONSISTENT (threshold = 5/8). The strict-rubric verdict is DEMOTE to EMERGING. The recommendation propagates to the v3.2.1 canonical card with evidence_tier_pre_phase7 = 'CONSISTENT' preserved.

The pattern. The H11.L13 demotion is not driven by any one gate. It is driven by four simultaneous gate failures (c2, c3, c4, c5) plus one UNVERIFIABLE (c6). Each of the four is the field-level deficit named in §8.2: c2 prospective-cohort funding gap, c3 ROBINS-I composite gap, c4 CI-reporting-standard gap, c5 cross-syndrome cohort-matching gap. The H11.L13 card is mechanistically supported by the field’s best vagal-inflammation literature and by Putrino’s autonomic group; the demotion is not a verdict on the biology. The demotion is a verdict on the field’s current capacity to report this biology in the forms peer-review-grade synthesis requires.

Recovery path. H11.L13 can recover ESTABLISHED tier without any new biological insight, simply by closing the four gates: (i) any prospective LC cohort that measures vagal-nerve TSPO-PET imaging or vagal-fiber-density biopsy outcomes at 12 + months credits c2; (ii) the same cohort’s ROBINS-I composite, scored at the recruitment-design level, credits c3; (iii) effect-size 95% CI reporting in any of the supporting citations credits c4; (iv) the ME/CFS vagal-neurotropism literature (Mackay 2021; Pollak 2023) catalogued and machine-extracted credits c5. The card has a clean recovery path with no biological discovery required — only reporting-form upgrades. This is the structural shape the four-gates field diagnosis takes at the per-card level: most demotions are reporting-form failures, not biology failures, and most have clean recovery paths through field-level reporting-standard or funding-design interventions.

§8.3 The four gates as the v3.3 research agenda

The four-chronic-gates pattern transforms what could be read as “atlas demoted 107 cards” into “Long COVID research design has four structural gates that need closing.” Each gate maps to a concrete v3.3 priority:

Gate	What it tests	Structural cause	v3.3 priority
c4	Machine-readable effect-size CI	No enforced 95% CI / null-clause reporting standard in LC journals	(a) Atlas schema: add effect_size_95ci; (b) Field: editorial reporting-standard adoption
c2	Prospective cohort or RCT design	Insufficient prospective-cohort funding ramp	(a) RECOVER-scale prospective cohorts extended to H5/H6/H10/H12/H13; (b) RCTs for the 5 intervention-deplete chains (§12.2)
c3	ROBINS-I composite ≤ moderate	No enforced ROBINS-I cohort tier; self-selected case-control dominant	(a) ROBINS-I-tier-enforced cohort recruitment standards; (b) Matched-deconditioning / matched-chronic-illness control cohorts at H6, H10
c1	PI-independence ≥ 2 groups	Single-PI subfield domination (Pretorius, Iwasaki, Putrino, Naviaux clusters)	NIH/Wellcome out-of-PI-group replication budget for top single-PI-led findings

§8.4 The atlas as honest grader of evidence forms, not of mechanism claims

A central feature of the four-gates finding is that the atlas does not claim the demoted 107 cards are wrong. The atlas claims that, under hostile peer-review standards, the cards cannot today be defended as ESTABLISHED. The honest tier for them is EMERGING (or for the 26 most-affected, SPECULATION) until the field publishes in forms that survive the four gates. The atlas is doing what GRADE methodology is supposed to do: it is grading the evidence, not the underlying claims, and the grades reflect the publication-standard gap.

A reviewer’s likely first attack on the four-gates finding will be: “if 96.3 % of demoted cards fail c4, isn’t the rubric too strict?” The atlas’s defense is that the rubric is correct, the field’s reporting standards are insufficient, and the dual-version v3.2 / v3.2.1 architecture is exactly the transparent representation of the gap (Supplementary Appendix B (relocated §7.7)). The conceptual canon v3.2 says “this mechanism is plausible and convergent across multiple observations”; the strict canon v3.2.1 says “this mechanism does not yet have the machine-readable evidence form that peer-review-grade synthesis requires”; both are true; both are valuable. The field-level fix is journal- and funder-driven; the per-paper fix is editorial; the atlas-level fix is schema-driven (v3.3 priority IM-12).

§8.5 v3.3 falsifier-pre-registration mandate

The c7 gate (named falsifier; 34.6 % not credited) is the smallest of the four chronic gates but the cheapest to fix structurally. The atlas’s falsification_trigger field was conceived as a narrative description; v3.3 should restructure it as {trial_or_experiment_type, registry_id_if_applicable, null_clause, post_hoc_falsifier_status}. The v3.3 falsifier-pre-registration mandate is: no card may carry tier > EMERGING unless its falsification_trigger names a concrete pre-registered trial or experiment with NCT identifier and null-clause specification. This converts the falsifier-discipline from a per-card hand-wave to a curation-process improvement, addressing 37 of the 107 demotions on a structural rather than per-card basis.

The falsifier-pre-registration mandate is also the structural reason behind the H3.L10 PR-1 disposition (§7.10): H3.L10 carries the largest BN counterfactual leverage in the atlas (+27.3 pp) but its falsifier (HEAL-LC NCT07597902) is only narratively encoded in v3.2; the strict rubric correctly demotes the card until the falsifier reads out and the post-hoc status is encoded.

§8.5.2 Structural barriers — why the field-wide pattern persists

The four chronic gates are not random methodological gaps. They are the downstream signature of three structural barriers in the LC research base that the field has not yet addressed. Naming the barriers explicitly makes the gate-closing program legible as field-level intervention rather than per-card complaint.

Barrier 1: No PROSPERO-equivalent falsifier registry. The PROSPERO international prospective register of systematic reviews is the gold standard for SR protocol pre-registration, and the GRADE working group’s protocols register prospective evidence reviews. There is no equivalent for mechanistic claims in LC or any other syndrome. A mechanism paper that proposes “sustained S1 antigenemia drives chronic neuroinflammation” does not pre-register the falsifier that would refute the claim. The Phase-7 c7 (named falsifier) gate captures this directly: 34.6% of demotion cards have no encoded falsifier. The structural fix is a per-mechanism falsifier registry analogous to PROSPERO — a public registry where mechanism papers pre-register the trial design, the null-clause boundary, the effect-size threshold for falsification, and the post-hoc-status outcome. The v3.3 priority IM-09 (“open atlas / open card extraction protocol with falsifier registry”) implements this at the curation layer; the field-level fix is editorial.

Barrier 2: No CI / null-clause reporting convention. The c4 (effect-size CI excluding null) gate fails at 96.3% across demotions. The structural cause is that LC mechanism papers report point estimates and p-values without bounding the 95% CI in machine-readable form. The fix is a journal-level reporting-standard adoption analogous to the CONSORT guidelines for RCTs or the STROBE checklist for observational studies. The fix is editorial-policy-level and has no per-paper or per-investigator cost — it is purely a publication-norm change. The atlas’s contribution is to demonstrate, quantitatively, how much of the LC mechanistic literature would gain Pearl-defensibility from this single editorial intervention.

Barrier 3: Prospective-design grant gap. The c2 (prospective design) gate fails at 81.3% across demotions. The LC literature is dominantly cross-sectional case-control because prospective cohorts are expensive (12+-month follow-up, biomarker measurement at enrolment plus longitudinal). NIH RECOVER and the German COVIM cohort are the two largest prospective LC cohorts globally; they cannot cover the entire mechanism stack. The structural fix is targeted prospective-cohort funding extended to neuroinflammation (H10), autoantibody (H5), mitochondrial (H6), HPA (H12), and pediatric (H13) subspecialty cohorts. The atlas’s intervention-deplete chain analysis (§6.4) is the quantitative map of where the prospective-design grant gap most hurts the field.

Barrier 4: Replication funding gap. The c1 (PI-independence) gate fails at 58.9% across demotions, dominated by UNVERIFIABLE non-credits where the atlas card has too few citations to determine PI-independence. The structural cause is single-PI domination of mechanistic subfields: Pretorius / Kell on microclots, Iwasaki / Akiko on autoantibodies, Putrino on dysautonomia, Naviaux on CDR. These groups produce excellent work; replicating their findings out of their groups requires independent funding that is not currently incentivized by NIH or equivalent funders. The fix is an explicit replication-budget line in LC funding portfolios — a structural intervention with a measurable, near-term impact on the field’s evidence-form profile.

These four structural barriers — falsifier registry, CI reporting convention, prospective-cohort funding, replication funding — together explain the four-chronic-gates pattern. None is intractable. Each is remediable with editorial, funder, and registry interventions at modest cost. The atlas’s contribution is to make the field-level barriers measurable and to name the specific interventions that would close them.

§8.6 The field diagnosis as a contribution

The four-chronic-gates finding is a contribution to the methodology of multi-mechanism mechanistic synthesis, not just to Long COVID. Any future mechanistic atlas — for ME/CFS, fibromyalgia, IBD, autoimmune-spectrum syndromes — will face the same four gates. The v3.2 atlas’s contribution is the structural identification of which gates fail at which frequencies, with the implication that future atlases should design ROBINS-I, prospective-cohort tracking, and machine-readable effect-size CI encoding into the curation pipeline from day one. Section §2.7 (methodological lessons for future atlases) develops this point in the manuscript’s discussion frame.

The field-diagnosis framing is also a defensive move: a peer reviewer who reads the four-gates finding as “the atlas is over-strict” can be answered with the dual-version architecture (the conceptual canon v3.2 preserves the mechanism-plausibility judgment; the strict canon v3.2.1 enforces the peer-review-defensible filter), the per-criterion breakdown (it is not the rubric’s strictness; it is the field’s reporting-standard gap), and the v3.3 priority queue (the field can close each gate with concrete editorial / funder / cohort-design moves). The framing converts what could be a defensive crouch into a constructive research-agenda contribution.

§8.6.2 The central methodological finding of v3.2

If the v3.2 atlas is read as a single-finding paper, the four-chronic-gates field diagnosis is the finding to take away. The atlas’s mechanism findings (the three-root architecture, the H7 ↔︎ H12 cycle, the H12 reclassification, the SIM01 keystone) are important and defensible, but they are mechanism-camp-specific findings that other syntheses will revisit and revise. The four-chronic-gates field diagnosis is portable: it applies, in principle, to any mechanistic synthesis of any chronic syndrome, and it provides the calibrated evidence-form mirror that the LC field has lacked and that ME/CFS, fibromyalgia, IBD, and chronic Lyme will need.

What the finding implies for journals. A journal that publishes LC mechanism papers can credit the c4 gate (effect-size CI excluding null) through editorial-policy intervention alone. The cost is zero per paper; the benefit is downstream Pearl-defensibility for every claim the paper supports. Reporting-standard adoption is the single most cost-effective field-level intervention identified by the atlas.

What the finding implies for funders. NIH, Wellcome, EU Horizon, and the Bill & Melinda Gates Foundation each have LC research portfolios. The atlas’s four-chronic-gates field diagnosis is a portfolio-allocation map: funding directed at the c2 gate (prospective cohort design), the c3 gate (ROBINS-I-grade recruitment), and the c1 gate (out-of-PI-group replication) closes the highest-frequency demotion drivers at the field level. The intervention-deplete chain map (§6.4) is the complementary RCT-priority map. Together they constitute a calibrated portfolio reallocation framework that funders can adopt without renegotiating their disease-area priorities.

What the finding implies for regulators. FDA and EMA review LC therapeutic claims under standard pivotal-trial criteria, but the LC research base’s evidence-form deficits (no enforced CI reporting; sparse prospective design; few replicated mechanism papers) make the regulator’s job structurally harder. The atlas’s evidence-tier mapping (ESTABLISHED → publication-grade; CONSISTENT → survival-grade; EMERGING → working-hypothesis; SPECULATION → candidate-hypothesis) is a regulator-legible framework for distinguishing which mechanistic claims warrant regulatory consideration in the context of a therapeutic submission and which do not. The framework is portable: any regulatory body assessing LC therapeutics can use it.

The four-chronic-gates finding is the central methodological finding of v3.2 and is the contribution we expect to outlast any specific mechanism finding in this manuscript. The mechanisms will be re-evaluated as new evidence lands; the field-level pattern they reveal will not change quickly, and the closing of the gates is a multi-year project. Naming the pattern is the contribution; closing the gates is the v3.3 and beyond agenda.

---

§9 Cross-syndrome non-specificity and cohort definitions

§9.1 Only 6 of 65 cells are LC_SPECIFIC

The T8 cross-domain coherence matrix tests each of the 13 chains against 5 comparator literatures for 65 chain-comparator cells. The verdict distribution is 19 VALIDATED / 19 PARTIAL / 21 INSUFFICIENT / 6 LC_SPECIFIC (§4.8; exhibit G29). Only 6 of 65 cells (9 %) are LC_SPECIFIC; 58 % VALIDATE or PARTIAL in at least one comparator. The 6 LC_SPECIFIC cells cluster in two places: R1 sustained S1 antigenemia + spike-protein persistence at 12+ months post-infection (paralleled in no comparator syndrome) and H11 meta-claim (asserted only in LC literature).

This is the empirical signature behind the cross-syndrome non-specificity finding (CF-3): LC is mechanistically continuous with the post-viral / autoimmune-syndrome spectrum on the chronic-tail maintenance layer, and LC-distinctive primarily on the upstream initiation layer (sustained R1 antigen). The hierarchical-DAG-with-annotated-cycles architecture (§4) handles this picture cleanly by separating the initiation layer (R1-dominated, COVID-specific) from the maintenance layer (mediator-cycle-rich, post-viral-shared).

§9.2 T10 negative-control battery: 7/13 partial-fail cross-syndrome

The T10 negative-control battery (§5.7; exhibit G30) finds 7/13 chains partial-fail the cross-syndrome NCE on the 12-month chronic-tail phenotype: R1, R3, H2, H5, H7, H10, H12. Only H6 passes outright (Appelman/Wüst 2024 PMID 38381103 activity-matched CPET-PEM design). Four chains (R2, H3, H8, H13) are untested at the NCE layer. H11 is structurally NC-paucity-bounded.

The cross-syndrome partial-fail signature is structurally consistent with CF-3 cross-syndrome non-specificity and with the §4.7 H11 biological-claim partial vindication. The signature does not mean the 7 chains are biologically null; it means that the chronic 12-month phenotype across these chains has substantial mechanistic overlap with post-viral / chronic-illness comparator syndromes, and that LC-specific external validation requires comparator-co-enrollment designs that the field has not yet run at scale.

§9.3 Cohort definition heterogeneity (PR-3)

Long COVID has at least four major cohort definitions in active use:

• WHO post-COVID-19 condition (PCC) — symptoms persisting ≥ 3 months after probable / confirmed COVID, with symptom duration ≥ 2 months not explained by alternative diagnosis.
• CDC PASC — broader definition encompassing any new, recurring, or persistent symptom > 4 weeks post-acute COVID.
• RECOVER PASC — research-cohort definition with prospective symptom-tracking and biomarker measurement.
• NICE NG188 — UK clinical-care definition with categorical symptom-cluster thresholds.

Cross-syndrome comparator literatures use yet more definitions: ME/CFS Fukuda / CCC / ICC / IOM criteria; POTS Heart Rhythm Society / Sheldon criteria; PTLDS IDSA / ILADS criteria. Many T10 partial_fail verdicts may be cohort-definition-driven rather than mechanism-driven. PR-3 disposition recommends adding a cohort_definition_dependence schema field per card — but Phase 7.5 explicitly defers this to v3.3 per the task spec’s scope limit. This is the largest single curation work item flagged but not addressed in v3.2.1.

The cohort-definition variance has two systematic implications for the atlas:

1. Comparator validation may underestimate cross-syndrome overlap — if ME/CFS cohorts use Fukuda criteria (broad) while LC cohorts use WHO PCC (narrower), the comparator may fail to validate a chain that would validate under matched criteria. The R1 antigenemia signature is the principal exception (Stein 2022 PMID 36517603 measures by direct autopsy, immune to cohort-definition variance).
2. Cross-syndrome partial-fail may be cohort-definition artefact — particularly for H12 (HPA axis, where Klein 2023 LC hypocortisolism PMID 37748514 differentiates LC from convalescent controls but is not yet cross-validated against ME/CFS-NIH cohorts), H7 (small-fiber neuropathy, where the PTLDS IENFD comparator literature uses different anatomical sites), and H13 (pediatric, where the comparator literatures are heterogeneously age-stratified).

The cohort-definition matrix is supplementary to this manuscript (manuscript-supplementary.md §S4) and is the v3.3 priority IM-13 (cohort definition tracking schema field).

§9.3.2 The PR-3 cohort-definition variance as a manuscript-level limitation

The cohort-definition variance flagged at §9.3 is not just a curation TODO. It is a substantive manuscript-level limitation that propagates through three layers of the v3.2 evidence stack. First, the T8 cross-domain coherence matrix’s per-chain VALIDATED / PARTIAL / INSUFFICIENT verdicts (§4.8) are conditional on the comparator-cohort definition that the cited papers used; a chain that PARTIAL-validates against an ME/CFS Fukuda cohort might VALIDATE against an ME/CFS CCC cohort and might INSUFFICIENT-validate against an ME/CFS IOM cohort. The atlas reports a single composite verdict per cell because the per-comparator-cohort granularity is not yet machine-extractable from the field’s primary evidence, but the composite verdict’s true uncertainty bound is wider than the verdict alone suggests. Second, the T10 negative-control NCE dimension (§9.2) inherits the same cohort-definition dependence: the seven chains that partial-fail NCE on the 12-month chronic-tail phenotype might pass under matched cohort definitions and might fail more decisively under stricter cohort definitions. Third, the SIM01 cross-domain VALIDATED verdict (§7) rests on the Giloteaux 2016 and Xiong 2023 ME/CFS gut-dysbiosis cohort papers, which use ME/CFS cohort definitions that differ from the WHO PCC definition used in Lau 2024; the cross-domain validation is mechanism-coherent but cohort-definition-loose. The honest reading is that the SIM01 cross-domain VALIDATED verdict survives reasonable cohort-definition perturbation, but the verdict is more robust under loose cohort matching than under strict cohort matching. The v3.3 atlas’s cohort_definition_dependence schema field (§12.3) is the structural fix; the v3.2 atlas reports the verdicts at the granularity the evidence supports and flags the propagation paths explicitly.

§9.3.3 ME/CFS as the principal comparator and what it tells us

ME/CFS is the principal cross-syndrome comparator for LC because its biology overlaps LC’s chronic-tail mediator stack most extensively, because its trial literature is more mature than LC’s at the cohort and biomarker level, and because the ME/CFS / LC clinical-phenotype distinction is the central practical question for downstream clinicians. The ME/CFS comparator literature validates seven of the LC chronic-tail mediator chains: R2 (Giloteaux 2016 PMID 28095889 and Xiong 2023 PMID 36450980 — gut dysbiosis; the SIM01 cross-domain anchor), H2 (multiple ME/CFS T-cell exhaustion cohorts), H5 (Müller 2024 ME/CFS GPCR-autoantibody cohort, parallel to Wallukat 2021 LC findings), H6 (Naviaux 2016 CDR / metabolomic ME/CFS signature, parallel to the Appelman / Wüst 2024 LC CPET-PEM design, and — added post-v3.2 closure by Phase 11f — Tasoula 2026 Front Immunol doi:10.3389/fimmu.2026.1776555, which performs the first direct same-tissue same-protocol transcriptional comparison of PCS-CFS vs T2bFA-CFS vastus lateralis biopsies and finds shared core programs of downregulated OXPHOS + elevated HIF signaling + mitochondrial stress across the two CFS phenotypes; this is the strongest peer-reviewed molecular evidence of PCS↔︎ME/CFS mitochondrial convergence in the published literature and directly addresses the §9.3.2 cohort-definition variance concern by holding tissue, methodology, and analytical pipeline constant across the comparator arms), H7 (multiple ME/CFS small-fiber-neuropathy cohorts), H8 (multiple ME/CFS mast-cell-activation cohorts), and H10 (Nakatomi 2014 PMID 24665088 ME/CFS PET-microglia neuroinflammation, parallel to the VanElzakker LC TSPO-PET findings; the synthesis-context preprint Martins 2025 doi:10.20944/preprints202506.1469.v1 reinforces the TSPO-PET, ASL-MRI, MR spectroscopy biomarker stack at the field-synthesis level but, as a non-peer-reviewed preprint, contributes no tier-rubric weight). Three LC chronic-tail chains have weaker or absent ME/CFS comparator coverage: H3 (the IFN-I-drive composite is more central to LC than to ME/CFS at the published-evidence level), H11 (the meta-claim is LC-literature-specific by construction), and H12 (the HPA-axis hypocortisolism finding is LC-specific in its strict form per Klein 2023 PMID 37748514, though older ME/CFS literature has parallel HPA findings). The ME/CFS comparator coverage validates the chronic-tail final-common-pathway interpretation of §9.4 and is the empirical basis for the recommendation (§12.14) that v3.3 invest in head-to-head LC vs ME/CFS prospective cohort designs that measure the same biomarker panel in both syndromes at matched time points. The expected outcome of such a design is that 60-80% of the chronic-tail mediator stack will overlap (the post-viral final-common-pathway prediction) and that LC will be distinguished by the sustained R1 antigenemia layer and the higher prevalence of acute-COVID respiratory/microvascular sequelae (the LC-distinctive layer). §9.4 prediction now has its first peer-reviewed quantitative molecular confirmation. Tasoula 2026’s PCS-CFS vs T2bFA-CFS direct comparison shows that the overlap at the level of core transcriptional programs (OXPHOS suppression, HIF signaling, mitochondrial-stress response) is more pronounced than at the pathway level — i.e., shared mitochondrial repression is accompanied by context-specific secondary remodeling programs (the PCS-CFS arm shows heightened ISR / UPR / innate-and-adaptive immune signaling that T2bFA-CFS does not). This is precisely the structure the §9.4 final-common-pathway interpretation predicts: shared upstream mitochondrial repression as the cross-syndrome final common pathway, with LC-specific antigen-persistence-driven immune amplification (R1 → H2 → H3) as the LC-distinctive overlay.

§9.4 Implication: most “LC-specific” mechanisms are post-viral final-common-pathways

The cross-syndrome non-specificity finding has a concrete clinical and research-agenda implication: most of the “LC-specific” mechanisms identified in the literature are actually shared post-viral / chronic-illness final-common-pathways. The chronic-tail mechanism stack (R2 dysbiosis, H2 T-cell exhaustion, H3 IFN-I drive, H5 autoantibody, H6 mitochondrial CDR, H7 small-fiber neuropathy, H8 mast-cell hyperactivation, H10 neuroinflammation, H12 HPA exhaustion) is concentrated in shared mediator nodes across ME/CFS, POTS, PTLDS, post-Ebola, and autoimmune-spectrum syndromes. The LC-distinctive feature is the sustained R1 antigenemia layer — sustained S1 spike protein, sgRNA detection, antigen reservoir in gut and tissue (Stein PMID 36517603, Goh PMID 34248921, Swank S1 longitudinal proteomics).

This implication is the empirical backbone of the §4.7 H11 biological-claim partial vindication: the field’s intuition that LC is multi-mechanism with feedback loops among mediators is empirically correct on the chronic-tail maintenance layer. The hierarchical model accommodates this through the H7 ↔︎ H12 cycle, the four asymmetric near-cycles, and the cross-syndrome-shared mediator pathways.

The implication for therapeutic identification (§6) is that interventions targeting the chronic-tail mediator pathways are likely to generalize across post-viral syndromes (SIM01-style synbiotic interventions targeting X.Dysbiosis are likely effective in ME/CFS as well; this is one reason the Lau 2024 ME/CFS gut-dysbiosis cross-validation is so structurally important). Interventions targeting the LC-distinctive R1 antigenemia layer (antivirals, anti-spike Ab) are likely LC-specific. The hierarchical model predicts the cross-syndrome generalization pattern.

§9.5 The sustained S1 antigenemia signature

Three anchor citations document the LC-distinctive sustained R1 antigen-persistence signature:

• Stein SR, Ramelli SC, Grazioli A, et al. SARS-CoV-2 infection and persistence in the human body and brain at autopsy. Nature. 2022;612(7941):758–763. PMID 36517603. Cross-tissue autopsy in 44 patients identified SARS-CoV-2 RNA persistence in 79 of 85 anatomical sites across multiple organs months after acute infection, including in patients with no detectable acute respiratory disease. This is the principal post-mortem evidence for sustained tissue reservoir.
• Goh D, Lim JCT, Fernández-Ruiz R, et al. Case report: Persistence of residual antigen and RNA of the SARS-CoV-2 virus in tissues of two patients with long COVID. Frontiers in Immunology. 2022;13:939989. PMID 36131932. Direct demonstration of persisting SARS-CoV-2 N-protein and RNA in gut mucosal biopsies of two long-COVID patients more than 200 days after acute infection, anchoring the gut as a candidate antigen reservoir compartment.
• Swank Z, Borberg E, Chen Y, et al. Measurement of circulating viral antigens post-SARS-CoV-2 infection in a multicohort study. Clinical Microbiology and Infection. 2024;30(12):1599–1605. PMID 39389851. Multicohort longitudinal proteomic measurement of circulating SARS-CoV-2 antigens (S1, full spike, nucleocapsid) using single-molecule array (Simoa) detection at picogram sensitivity, finding sustained S1 antigenemia in a substantial fraction of long-COVID patients at 14+ months post-infection in two of four cohorts, and providing the principal serum-based anchor for the sustained R1 antigen signature.

These three anchors form an internally coherent triad. Stein measures the tissue reservoir directly at autopsy (the unambiguous existence proof for sustained SARS-CoV-2 nucleic acid persistence). Goh localizes the antigen reservoir to gut mucosa in living long-COVID patients more than six months post-acute (the principal living-patient anchor for the R1 ↔︎ R2 mechanistic linkage that drives the SIM01 → X.Dysbiosis Pearl-defensible target). Swank measures the resulting circulating antigen burden in serum at 14+ months by orthogonal proteomic methodology in independent cohorts, closing the methodological triangle (autopsy → biopsy → serum). The triad satisfies the Bradford-Hill consistency criterion across three independent measurement modalities and three independent cohorts. The sustained S1 antigenemia signature is the principal LC-distinctive feature in the cross-syndrome non-specificity matrix (§9.1; 2 of 6 LC_SPECIFIC cells in T8) and the empirical answer to “what makes Long COVID a coherent diagnostic entity at all, given that 58 % of mechanistic cells validate or partially validate in comparator syndromes?”

The clinical implication of the signature is that R1 should be assayable as a peripheral biomarker for LC stratification. Swank’s Simoa platform is the candidate methodology. A v3.3 priority item flagged in §12 is to convert the antigenemia signature from a mechanistic finding into a deployable trial-enrollment biomarker: the most powerful R1-targeted RCT design would enroll only Simoa-S1-positive patients, eliminating the dilution that probably damaged the Geng / RECOVER-VITAL / Bonilla / STOP-PASC nirmatrelvir-treatment trials whose ITT populations were almost certainly heterogeneous on the antigen-positive / antigen-negative split (§11 develops this point). The atlas does not adjudicate whether antigenemia-positive LC is a clinically distinct phenotype with a distinct treatment response (the evidence to do that has not been generated); it points to the trial design that would generate the evidence.

A residual honest caveat is that the Goh PMID was upstream miscited in v3.2 source notes as PMID 34248921 (which is actually the Proal 2021 review Long COVID or Post-acute Sequelae of COVID-19 (PASC): An Overview); the correct PMID per the Frontiers in Immunology case-report DOI is PMID 36131932, verified against analysis/bibliography/bibliography-master-v3.2.json. The miscitation propagated through several Phase 3–6 working memos before being caught by the §10 bibliography sweep; the canonical reference in this manuscript is the verified Frontiers in Immunology PMID 36131932. The 4 Goh-anchor downstream claims are mechanistically unchanged. The miscitation is logged in §10.5 and in the Phase 6 T11 bibliography audit memo (analysis/bibliography/resolution-log.md).

---

§10 Limitations

The v3.2 atlas earns its conclusions only insofar as the limitations of the synthesis are stated as plainly as the findings. Eleven analytical techniques converge on a measured picture; the convergence is meaningful precisely because the failure modes are catalogued. This section enumerates the twelve principal limitations of the v3.2 architecture, ordered roughly from most-to-least structural.

§10.1 Single-LLM review (PRISMA 2020 item 9)

The atlas was curated and audited by a single LLM-driven review pipeline (Claude 4.7-Opus, supervised by a single Captain). PRISMA 2020 item 9 requires that “two or more reviewers screened each record and each report retrieved.” The atlas does not satisfy this requirement. The Phase 6 T6 PRISMA audit (analysis/prisma-audit/prisma-2020-fleet-summary.json) records the dual-reviewer item as FAIL across all 13 chains. This is the single most consequential PRISMA failure for the per-card evidence-grading layer.

The atlas mitigates this with three structural compensations: (1) adversarial red-team review at Phase 6 (CN-1 through CN-9 challenge nodes, exhibit G24), (2) Phase 7 strict re-tier with an independent eight-criterion rubric that demoted 107 of 222 cards including many top-tier cards, and (3) per-card audit trail with explicit phase7_audit_status annotation in v3.2.1. These compensations are not equivalent to dual-reviewer per PRISMA and the manuscript does not claim otherwise. v3.3 should add a second independent LLM reviewer to the per-card screen, ideally from an independent model family (e.g., Gemini 2.5 or GPT-5), with disagreement adjudication via Captain or a third agent.

§10.2 Eight PRISMA structural items fail across all 13 chains

The Phase 6 T6 PRISMA 2020 audit returns eight items that fail across all 13 chains: 9 (dual reviewer, §10.1 above), 14 (effect-measure pooling — atlas does not pool effects across heterogeneous trial designs; it ranks mechanism evidence per card), 15 (synthesis methods specification beyond pooling), 17 (reporting bias / publication-bias assessment per outcome — atlas does some via T6 funnel-plot analysis where ≥ 10 RCTs exist per outcome, but most cards do not meet the threshold), 18 (certainty-of-evidence assessment per outcome via GRADE — atlas implements a GRADE-equivalent five-tier vocabulary but not the per-outcome GRADE summary table), 22 (results of individual studies — atlas references per-card evidence but does not report extracted effect sizes per study), 23 (results of syntheses — atlas reports rank distributions and chain-vulnerability scores but not pooled effect estimates), and 27 (registration and protocol — the atlas pre-registration docs/v3.2-PRD.md is internal-not-PROSPERO-registered).

These items fail correctly. The atlas was not built as a PRISMA-compliant systematic review and does not claim to be one. PRISMA is the appropriate audit standard for trial-effect-size syntheses; the atlas is a mechanism-rank synthesis layer over the trial corpus. The eight FAIL items are reported here so that downstream readers can assess the gap between the atlas and a putative future LC SR (the gap is large but well-specified). The §2 conclusion frames this as “the atlas is a complement to, not a replacement for, future PRISMA-compliant LC systematic reviews.”

§10.3 Fifty cards carry HIGH_UNVERIFIABILITY in Phase 7 audit

The Phase 7 strict re-tier flagged 50 of 222 cards (22.5 %) with phase7_audit_status = 'HOLD' and unverifiability_flag = 'HIGH'. The flag means that one or more of the eight strict criteria (C1 primary evidence specificity, C2 effect-size CI completeness, C3 prospective-cohort tracking, C4 quantitative parameter granularity, C5 mechanism-pathway specification, C6 dose-response when applicable, C7 named falsifier with NCT or experiment, C8 cross-comparator validation) could not be confidently graded from the card’s encoded primary-evidence references. The HOLD verdict is a Phase 7 mechanical artefact and does not assert that the underlying biological mechanism is null; it asserts that the card’s machine-readable schema does not support strict-criteria adjudication.

The 50 HOLD cards are distributed approximately uniformly across chains (4–6 per chain) and concentrate in H7 (small-fiber neuropathy, where extractable CI fields are field-typically absent), H8 (mast-cell, where falsifier specification is field-typically narrative), H12 (HPA axis, where prospective cohort tracking is field-typically thin), and H11 (meta-claim, where C1 primary-evidence-specificity is structurally underspecified). The v3.3 priority queue (§12) earmarks Phase 8 adjudication of the 50 HOLD cards as ~ 8 hours of focused Captain-supervised review; the expected outcome is roughly 25 cards re-tier-up under strict criteria, 15 confirm as HOLD on substantive grounds, and 10 demote under strict criteria. None of the §4, §5, §6, §7, Supplementary Appendix B headline findings depends on the resolution of the 50 HOLD cards.

§10.4 The four-chronic-gates failure rates reflect field deficits, not atlas weakness

The four chronic gates (C4 quantitative parameters 96.3 %, C2 effect-size CI 81.3 %, C3 prospective cohort 72.9 %, C1 primary evidence specificity 58.9 %) fail at frequencies that one might prima facie attribute to atlas over-strictness. The §8 field diagnosis argues that these failure rates are structural features of the LC literature: most LC primary-evidence papers do not report quantitative parameters at the granularity strict GRADE requires, do not provide confidence intervals on extractable effect sizes, do not implement prospective cohort tracking beyond 12 months, and rely on retrospective cohort or cross-sectional designs that fail the strict C1 gate.

The limitation is not symmetric. It is not the case that the atlas would pass if curated by a different reviewer; it is the case that the LC literature would not pass any strict eight-criterion rubric, and the atlas is faithfully reporting that fact. This is a limitation in the sense that the manuscript’s per-card evidence-tier distribution (especially the 13.5 % ESTABLISHED rate after strict re-tier) is field-deficit-limited rather than atlas-method-limited; readers should interpret the strict-rubric demotions as what the field has not yet produced, not as what the atlas has failed to find. v3.3 will track field improvements via re-audit at 18-month intervals; the §12 priority queue includes specific editorial / funder / cohort-design recommendations.

§10.4.2 Phase 11f post-closure addendum scope and limits

The v3.2 atlas is the immutable canonical release as of v3.2.1 strict-rubric closure on 2026-06-04. Two peer-reviewed-or-preprint papers — Tasoula et al. 2026 Front Immunol (peer-reviewed, doi:10.3389/fimmu.2026.1776555, published 21 May 2026) and Martins et al. 2025 Preprints.org (not yet peer-reviewed, doi:10.20944/preprints202506.1469.v1, posted 17 June 2025) — were integrated post-closure as the Phase 11f addendum. The integration is bounded by the Phase 7 strict rubric. Tasoula 2026 strengthens H6 within the ESTABLISHED tier (the chain already carries 7 ESTABLISHED-tier cards — tied for the largest survivor count) by adding a peer-reviewed multi-omics anchor that contributes to c5 cross-domain coherence (the direct PCS-CFS vs T2bFA-CFS comparison moves the H6 ↔︎ ME/CFS T8 verdict from VALIDATED-by-comparator-citation to VALIDATED-by-direct-molecular-convergence) and c6 negative-control reasoning (multi-omics, cross-species, cross-tissue replication of the OXPHOS-suppression signature) but does not contribute c2 prospective-design weight (the Tasoula analysis is a re-analysis of previously published transcriptomic, proteomic, and metabolomic datasets and is appropriately observational in design). The Phase 11f addendum therefore does not re-tier H6; it strengthens the chain within the ESTABLISHED tier and is documented at this granularity in the manuscript-expansion-changelog and in NATURE-AI-COMPLIANCE.md. Martins 2025, as a preprint synthesis, cannot contribute to chain-tier promotion at all (fails c2 prospective-design and c8 peer-review-equivalent gates under the strict rubric); it is cited only as a synthesis-context anchor for the H10 neuroinflammation chain and the candidate biomarker stack (GFAP, sTREM2, S100β, IL-6, TNF-α, TSPO-PET, ASL-MRI, MR spectroscopy, diffusion MRI), with every Martins citation in the body flagged as preprint per the §3.10 honest-gap disclosure protocol. Novel mechanistic candidates from Martins (local sleep intrusions, impaired memory reconsolidation, astrocyte-mediated network destabilization) are pre-registered as v3.3 hypothesis-generating items in §12.4, not as v3.2 evidence-promotion entries. The addendum partially resolves prior limitations around cross-cohort molecular comparison (Tasoula’s same-tissue same-protocol design) and longitudinal molecular evidence (Tasoula’s 12-month PBMC and 1-and-6-month proteomic horizons) but does not resolve the dual-reviewer (§10.1), prospective-RCT (§10.10), or pediatric-cohort (§10.12) gaps.

§10.5 Bibliography sweep flagged 8 PMIDs as unverified

The Phase 6 T11 bibliography audit (analysis/bibliography/resolution-log.md, §T11.8) flagged eight in-text PMIDs as unverifiable against PubMed, Europe PMC, Crossref, and OpenAlex APIs at the audit timestamp: three were typographical errors (single-digit transpositions in source memos), three were superseded by version-of-record DOI updates (preprint → published-journal PMID migration), one was a category mismatch (the cited PMID resolved to a different author’s paper on a similar topic — the §9.5 Goh case, where PMID 34248921 was upstream-miscited and resolves to Proal 2021 rather than the intended Goh 2022 PMID 36131932), and one was a withdrawn preprint not yet PMID-assigned. The eight cases are catalogued in analysis/bibliography/unresolved.md. The §14 reference list in this manuscript uses only verified PMIDs; in-text references that are not yet verified (or that propagated through Phase 3–6 memos under a now-corrected PMID) are marked [PMID PENDING VERIFICATION] per the §3.10 honest-gap disclosure protocol.

§10.6 T2 MR ABO: single-SNP β=+0.27 p=0.027 is the canonical citation

The T2 Mendelian-randomization layer (§5.4) tested four published GWAS-derived instruments against three LC outcomes. The ABO blood-group instrument is the most consequential because ABO is the only locus with published-GWAS-confirmed reach into LC severity (Lammi 2025 PMID 40399555). The aggregated 3-SNP IVW analysis returns a non-significant effect (p ≈ 0.13), but inspection of the SNP-level Wald ratios reveals that one of the three SNPs is a known horizontal-pleiotropy instrument for ABO ↔︎ VWF / FVIII (which themselves affect microclot kinetics and confound R3 estimation). The empirical correction is to retain the single-SNP Wald estimate from the non-pleiotropic ABO instrument: β = +0.27, p = 0.027 NOMINAL. The aggregated 3-SNP IVW result is preserved in the methodology supplement (manuscript-supplementary.md §S2.3) for full transparency.

This means the canonical T2 MR ABO citation in the manuscript main text is the single-SNP nominal-significance result, not the aggregated null result. The atlas does not claim genome-wide significance for ABO → LC severity; it claims a nominal-significance directional signal that is biologically coherent with the R3 microclot mechanism and the Lammi GWAS finding. Readers who prefer the aggregated 3-SNP result should consult the supplement; the choice between aggregations is methodologically substantive and is disclosed here so that downstream meta-analyses can recompute under either convention.

§10.7 T1 BN-vs-T12 disagreements: cycle-broken edges H7→H8, H7→H12

The T1 Bayesian-network ranking (analysis/bayesian-net/run_bayesian_net.py, §5.1) and the T12 cycle-graph ranking (analysis/cycle-graph/, §5.5) disagree on two edge orientations: H7→H8 (T1 ranks the edge directional with H7 upstream; T12 finds a near-cycle and arbitrates H7↔︎H8 as bidirectional with weak primary direction H7→H8) and H7→H12 (T1 finds a directed edge; T12 finds the explicit H7↔︎H12 strong cycle reported in §4.5). T12 wins arbitration in both cases because the cycle-graph analysis incorporates the temporal-precedence evidence that the BN’s static DAG cannot represent. The T1 outputs are preserved in the methodology supplement; the §4 architecture results, §5 sensitivity layer, and §6 therapeutic identification all use the T12 cycle-graph as canonical.

§10.8 Dual-version architecture: v3.2 conceptual canon vs v3.2.1 strict canon

The v3.2 / v3.2.1 dual-version architecture (Supplementary Appendix B) means that the manuscript cites two parallel canonical versions: architecture, cycle structure, and inter-root verdicts cite v3.2 (the immutable conceptual canon that locks the 13-chain partition, the 27 mediator nodes, the H7 ↔︎ H12 cycle, and the four asymmetric near-cycles); per-card evidence-tier claims and the strict-criteria audit cite v3.2.1 (the strict-rubric-compliant canon that demoted 107 cards including the H3.L10 HEAL-LC anchor pending NCT07597902 readout). The dual citation policy is documented in Supplementary Appendix B (relocated §7.7) and is the structural reason why a reader cross-checking a card’s tier in v3.2 source vs. the manuscript-quoted tier may find a one-tier discrepancy: the manuscript reports v3.2.1 strict-canon tiers in §6–§8 unless the surrounding paragraph specifies v3.2 conceptual-canon tier. This is a documentation limitation (the citation policy is implicit in section context rather than per-claim tagged) that v3.3 will address by adding explicit [v3.2] / [v3.2.1] provenance tags per per-card claim.

§10.9 T7 NMA topology has zero closed loops — no consistency test possible

The T7 network meta-analysis catalogues 42 interventions across the LC corpus and finds that the network topology is a star around placebo with nine placebo-bridged edges and zero head-to-head intervention-vs-intervention trials. The structural consequence is that no closed loops exist, no Bucher-Higgins-Welton consistency test is computable, and no defensible SUCRA or P-score across the catalogued interventions can be reported. The atlas explicitly refuses to publish a SUCRA ranking, which is the right epistemic move under the field’s current trial-corpus topology but is a substantive limitation on the manuscript’s therapeutic-prioritization claims.

The limitation also constrains the per-class effect estimate reporting in exhibit G34: each class’s summary estimate is anchored to placebo-controlled contrasts only, with the explicit caveat that across-class pooling is not defensible. A reader who wants a quantitative “which class is best” estimate cannot obtain one from this manuscript; the right reading is that SIM01 (single trial, single class) is the Pearl-defensible therapeutic claim under the T9 framework (§7), and the other classes are not yet at the trial-corpus topology to support pooled ranking. The closing of this gap requires investment in head-to-head trial designs, which is structurally non-trivial in LC (high cost, slow recruitment, mechanism-camp-specific cohort stratification challenges).

§10.10 T9 counterfactual simulation is model-based, not observed

The Pearl do-calculus counterfactual simulation (T9, §6.2) computes intervention-priority scores P = E × I × V under the assumption that the T1 Bayesian network’s edge weights and posterior distributions are sufficient to estimate the counterfactual posterior shift do(X). This is a model-based assumption, not an observation: the T9 effect-weight column E is computed from the T1 posterior, not from an observed do() intervention. The only intervention where the assumption is validated against an observed RCT is SIM01 (Lau 2024 PMID 38071990), where the trial randomisation provides the back-door identification directly and the observed primary-endpoint effect matches the T9 model-predicted relief on the GI / fatigue / brain-fog / concentration symptom domains in direction and rough magnitude.

The nine remaining T9 candidates’ priority scores are model-only and inherit the T1 posterior’s uncertainty plus the cycle-iteration sensitivity bound (§6.2). The honest interpretation is that the T9 ranking is the best model-based intervention prioritization the v3.2 evidence supports, not the observed effect ranking. A reader should treat T9 P = 9.00 for SIM01 as observation-validated and T9 P scores < 9.00 for the other nine as model-based predictions awaiting RCT confirmation. The v3.3 priority queue (§12) identifies the specific RCTs that would convert each model-based ranking into observation-validated.

§10.10.2 T10 negative-control battery sparse on within-LC tracer NCs

The Lipsitch / Tchetgen-Tchetgen negative-control battery (T10, §5.7) implements three NC categories: cross-syndrome NCE (does the chain replicate in comparator syndromes? — 7/13 chains partial-fail on the 12-month chronic-tail phenotype), within-LC tracer (is the within-LC observation pattern consistent with the chain biology? — H6 passes outright on the activity-matched CPET-PEM design), and NCO (is an unrelated outcome unaffected by the same exposure? — sparse coverage in the LC corpus). The within-LC tracer dimension is where the LC literature is thinnest: most LC cohort studies measure symptom outcomes but few measure both the asserted-causal biomarker and an unrelated tracer outcome that should be unaffected if the causal claim is true.

The T10 verdict map (exhibit G30) is partly UNVERIFIABLE on the NCO dimension across many chains. The atlas grades T10 conservatively and credits c6 (negative control) only when at least the NCE or tracer dimension returns PASS or PARTIAL; cards where T10 returns structural NC paucity are flagged UNVERIFIABLE and do not credit c6. This is the strict-rubric-correct reading but it understates the chain biology’s actual negative-control profile in cases where the field has not yet measured the right outcome pair. The v3.3 priority queue includes targeted NCO measurement on the H7, H10, and H12 chains where the within-LC tracer dimension is the v3.2 gap.

§10.11 Phase-7 rubric over-harshness on single-high-quality-study cards

The strict eight-criterion rubric is designed to be hostile-reviewer-safe. The cost of that design is that a card whose evidence rests on a single very-high-quality study (e.g., a single large prospective RCT with low ROBINS-I, machine-readable CI, named falsifier, and sign consistency) cannot earn c1 PI-independence credit even if the single study is methodologically perfect. The c1 gate requires ≥ 2 distinct PI groups demonstrated; a single excellent paper from a single group fails this gate by construction. The atlas implements this rule deliberately because PI-independence is a structural anti-confound against publication-bias, p-hacking, and single-laboratory artefacts. The cost is that a small number of v3.2 cards (the atlas estimates ≈ 8 cards across the corpus) demote to CONSISTENT or EMERGING under c1 failure when the underlying evidence is, at the per-paper level, ESTABLISHED-tier.

The Appelman 2024 PMID 38177128 activity-matched CPET-PEM design is the most-affected example: a methodologically outstanding single paper from a single group that does not yet have an independent replication. The atlas’s c1 disposition is to demote the relevant H6 cards to CONSISTENT rather than ESTABLISHED until an independent replication lands. This is the strict-rubric-correct disposition and is a feature, not a bug, of the eight-criterion design. Readers who weight individual-study quality more heavily than PI-independence will reasonably disagree with specific demotions; the dual-version v3.2 / v3.2.1 architecture (Supplementary Appendix B) is the transparent representation of the resulting tier gap.

§10.12 H13 pediatric: 98 unscored PMIDs, 0% ROBINS-I coverage, 0 ESTABLISHED-tier survivors

The H13 pediatric LC chain carries the highest unscored-PMID burden in the atlas (98 PMIDs) and the lowest ROBINS-I coverage (0%). The Phase-7 audit finds 0 ESTABLISHED-tier survivors in H13. This is a structural limitation of the v3.2 atlas: the pediatric LC literature is rapidly developing and the v3.2 sweep landed at a corpus state where the available evidence does not yet support strict-rubric ESTABLISHED tier for any pediatric mechanism card. The limitation is most consequential because pediatric LC trial design is structurally distinct from adult (developmental trajectory effects, MIS-C cross-talk, regulatory consent constraints) and the v3.2 atlas’s adult-mechanism findings do not transfer to pediatric LC without explicit cohort-design adaptation.

The v3.3 priority IM-13 (pediatric H13 prospective cohort + RoB pass) is the highest-priority single chain refresh in the atlas’s research agenda. Until that refresh lands, the manuscript’s pediatric LC claims are limited to the cross-domain inference layer: the adult-mechanism stack is mechanistically continuous with pediatric LC at the H2, H5, H7, H8, H10 chain level, but the chain-specific ESTABLISHED-tier evidence does not yet exist in pediatric cohorts.

§10.12.2 Case-series exclusion and the conservative-rubric trade-off (Pridgen evidence package)

The atlas’s strict Pearl-rubric and RCT-anchored evidence-tier framework is designed to maximise specificity at the cost of sensitivity. The trade-off has a known cost: a large body of case-series-grade clinical experience accumulated outside the RCT corpus is systematically under-weighted by the rubric, and that under-weighting is most consequential for a disease (Long COVID) that as of v3.2 has zero FDA-approved treatments and a small completed-RCT corpus. The canonical example surfaced by Dr. Skip Pridgen’s evidence package (ingested into v3.2 at the Phase 11d boundary) is the IMC-1 / IMC-2 ± Paxlovid clinical experience: a decade of off-label use across thousands of patients in the Pridgen clinic, recently formalised in an IRB-approved prospective open-label LC case series (Pridgen et al. 2026, PMID 41562079) reporting a Cohen’s d ≈ 1.8 effect on PGIC fatigue with 731-day durability — evidence that the rubric tier-assignment vocabulary cannot promote past PARTIAL/CONTESTED-POSITIVE because it is not RCT-grade in LC, but evidence that any clinician triaging the chronic-tail LC population would treat as substantively informative. The Phase 11d ingestion re-scores H3.L10 from P = 1.28 to P ≈ 3.50 to reflect the broader IMC-2+Paxlovid combination protocol (§6.2) and adds the explicit FORTRESS-monotherapy-vs-combination distinction to §6.4.2, but the underlying methodological tension persists: the rubric is calibrated for what survives strict Pearl-defensibility, not for what is clinically valid. The Phase 11e Pridgen personal-communication addendum (2026-06-05) bumps H3.L10 a further notch to P ≈ 3.75 to credit the FORTRESS cohort-2 (post-vaccine) disclosure that the cohort met both primary and secondary endpoints; this is itself a paradigm instance of the case-series-tier evidence-vocabulary gap this section identifies. A personal-communication trial-cohort-disaggregation disclosure from a trial PI is informative under any reasonable clinical-decision-making rubric, but the v3.2 rubric has no explicit tier name for it; v3.3’s case-series-tier vocabulary should explicitly cover “personal-communication trial-cohort-disaggregation disclosures” as one of its sub-types, with a published cohort-stratified reanalysis as the natural falsifier/ratifier.

The atlas flags three structural mitigations for the v3.3 refresh: (a) an explicit “case-series-grade clinical experience” tier in the per-card evidence vocabulary, sitting below CONSISTENT but above SPECULATION and carrying its own falsifier-pre-registration requirement (a planned RCT NCT identifier that would either ratify or falsify the case-series signal); (b) a per-chain “strict-rubric / clinical-experience” delta column in the v3.3 priority queue exhibit, making the conservative-rubric cost legible per row; (c) explicit pre-registration of the Pridgen-protocol HEAL-LC NCT07597902 readout as the v3.3 trigger for re-tiering H3.L10. None of these is a retraction of the strict-rubric posture; all are explicit-trade-off acknowledgments that the field’s evidence-generation infrastructure has not yet produced the RCT-grade replications the rubric is calibrated against. Captain Pridgen’s first objection — excluding a large body of case-series-grade clinical experience for a disease with no FDA-approved treatments is methodologically conservative but practically problematic — is logged here as a known limitation rather than as a falsification of the v3.2 atlas’s conclusions.

§10.13 Chain-vulnerability scoring is LLM-panelist-derived

The T5 chain-vulnerability score (§5.5) is computed from a simulated three-LLM-panelist hostile peer review under five weighted criteria. The scoring is LLM-panelist-derived and inherits the same architectural limitation as the ROBINS-I scoring (§10.1): it is not a gold-standard human peer-review pass. The chain-vulnerability scores are reported as comparative rankings (H6 = 8.5, H10 = 8.0, R1 = 7.5, H11 = 3.0) where the ordering is robust under reasonable panelist variation, not as absolute peer-review-rejection-probability estimates. A future v3.3 should instrument the T5 layer with at least one independent human-panelist pass against a stratified sample (≈ 6 chains) to validate the comparative ordering.

§10.14 Atlas at a glance (G39)

The atlas-at-a-glance summary panel (exhibit G39) compresses the full v3.2 deliverable into a single page: 5,101 papers, 222 link cards, 13 chains, 27 mediators, 117 polytope cells, 156 inter-root pair verdicts, 4 SPOFs, 30 ESTABLISHED-tier survivors, 1 Pearl-defensible therapeutic claim, 4 chronic field-level gates. The panel is designed for landing-page or abstract-graphic deployment and is the appropriate visual handoff to clinicians, funders, and journal editors who want the headline picture without traversing the full manuscript. The panel is explicitly identified as a summary, not a substitute for the per-chain evidence: a reader who acts on the panel without consulting the underlying chain narratives is reading the atlas at the wrong granularity. The panel’s role is to invite the reader into the manuscript, not to replace it. The full caption is reproduced in manuscript-exhibits-table.md.

---

§11 Discussion

The eleven analytical techniques applied to 5,101 papers and 222 mechanism link cards return a measured, internally coherent, and structurally surprising picture of Long COVID’s mechanistic architecture. The discussion organizes the principal findings around five themes: (1) R1 viral persistence as principal but not sole driver, (2) the H11 two-frame synthesis, (3) SIM01 → X.Dysbiosis as the sole Pearl-defensible therapeutic claim, (4) the four single-point-of-failure cards and five RCT-deplete chains, and (5) the atlas-vs-Groysman synthesis. We close with anticipated reviewer challenges.

§11.1 R1 viral persistence is the principal driver, not the sole driver

The §4.1 three-root architecture identifies R1 (sustained viral antigen / SARS-CoV-2 persistence) as the principal upstream driver, with strong PROBABLE_CAUSE outgoing reach into nine of the ten H-chains. This finding does not mean that eliminating R1 cures LC. The Paxlovid prevention / treatment asymmetry is the most consequential empirical signal the atlas integrates and the strongest single argument against monocausal R1 framing.

Three large-N nirmatrelvir-during-acute-COVID studies report positive prevention effects: Xie 2023 PMID 36951829 (US VA cohort, hazard reduction for LC outcome ≈ 26 %), Bramante 2023 PMID 37302406 (COVID-OUT RCT secondary outcome, prevention of post-COVID condition over 10 months), and Ioannou 2023 PMID 37903369 (VA cohort, propensity-matched prevention of post-COVID outcomes). At least four nirmatrelvir-during-established-LC trials report null treatment effects: Geng 2024 PMID 38848477 (STOP-PASC at Stanford, primary symptom endpoint null), RECOVER-VITAL (interim readouts to date null on the primary composite), Bonilla 2023 PMID 36969241 (RECOVER-Adult interventions taskforce summary listing nirmatrelvir as null on established LC), and a fourth phase-2 trial currently in readout window. The asymmetry is empirically robust.

The biological interpretation that the atlas favors is that R1 is necessary for LC initiation but not sufficient for LC maintenance once the chronic-tail mediator cycles (R2 dysbiosis, H2 T-cell exhaustion, H3 IFN-I drive, H5 autoantibody, H6 mitochondrial CDR, H7 small-fiber neuropathy, H8 mast-cell, H10 neuroinflammation, H12 HPA exhaustion) are established. The chronic-tail cycles are mediator-driven and self-sustaining; eliminating the upstream antigen reservoir at 12+ months does not necessarily collapse the downstream cycles. This interpretation is [CONSISTENT] with the §4.5 H7 ↔︎ H12 cycle structure, the §5.7 T10 cross-syndrome NCE partial-fail signature (7 of 13 chains fail the cross-syndrome NCE on chronic-tail phenotype), and the §9.4 cross-syndrome shared-final-common-pathway implication. The interpretation is [NOT ESTABLISHED] in the sense that no published trial has yet tested the strong form (early R1 elimination prevents LC; late R1 elimination does not reverse LC) under proper sub-group analysis on antigenemia-positive vs antigenemia-negative LC.

The atlas’s specific R1-related research-agenda recommendation (§12) is the antigenemia-biomarker-stratified treatment RCT: enroll only Simoa-S1-positive LC patients (using the Swank PMID 39389851 methodology), randomize to nirmatrelvir vs placebo for 30 days, and measure both R1 antigen clearance and clinical-response on the standard LC symptom composite. This trial has not been run. It is the single most-informative R1-targeted trial design the atlas can propose.

§11.2 The H11 two-frame synthesis

The cross-root amplification cascade that the v3.2 architecture decomposes is shown in exhibit G54 — a three-column flow diagram from the three roots through the twenty-seven mediators to the nine symptom domains, with edge thickness encoding the T1 Bayesian posterior amplification weight. The cascade visualization makes the H11 two-frame synthesis concrete: every flat-network connection that the Groysman frame asserts is rendered as an edge in G54, but the hierarchical-DAG layer ordering is preserved so the initiation → amplification asymmetry remains visible. Exhibit G54: Cross-root amplification cascade from the 3 causal roots through the 27 mediators to the 9 symptom domains. Edge thickness encodes the T1 Bayesian-network posterior amplification weight. Top-4 mediators (X.Dysbiosis 0.234, X.IFN-I 0.187, X.T-cell exhaustion 0.162, X.Mitochondrial dysfunction 0.142) absorb ~58% of downstream symptom variance. The H7⇌H12 cycle is highlighted in red; the SIM01 keystone edge (R2 → X.Dysbiosis) is highlighted in gold. The diagram concretizes the §11.2 two-frame synthesis between the Groysman flat-network and the hierarchical-DAG-with-cycles architecture.

The published H11 framework (Groysman 2024, multi-mechanism flat-network with feedback loops as the meta-claim) fails the C1–C8 strict eight-criterion rubric universally (T5 vulnerability score 3.0; tied for highest in atlas) on grounds that the meta-claim is not derivable from a single primary-evidence study, is asserted only in the LC literature (T8 H11 row LC_SPECIFIC), and does not specify a falsifier. The published Groysman framework as a peer-review-defensible claim is [NULL].

The biological claim behind Groysman — that LC is multi-mechanism with feedback loops among R1, R2, H2–H13 mediators rather than dominated by any single driver — is [PARTIAL VINDICATION] under the atlas’s stress tests. Four independent layers of evidence point at the multi-mechanism-with-cycles structure: (1) T12 cycle-graph analysis identifies the H7 ↔︎ H12 strong cycle and four asymmetric near-cycles (§4.5); (2) T7 NMA topology analysis (§6.1) finds zero closed loops in the trial network (consistent with the field having tested individual mechanisms in isolation and not having tested the multi-mechanism combinations that the Groysman frame would predict); (3) T8 cross-domain coherence analysis (§5.6) finds 58 % of mechanism cells VALIDATE or PARTIAL in ME/CFS / POTS / PTLDS / post-Ebola / autoimmune-spectrum comparator literatures (the chronic-tail mediator stack is shared across post-viral syndromes); and (4) T10 cross-syndrome NCE battery (§5.7) finds 7 of 13 chains partial-fail the cross-syndrome NCE, indicating that the LC chronic-tail phenotype is mediator-cycle-dominated rather than R1-dominated at 12+ months.

The hierarchical-DAG-with-annotated-cycles architecture (§4) absorbs both frames cleanly: the v3.2 atlas asserts that LC has a directed initiation layer (R1 dominant, COVID-specific) and a chronic-tail maintenance layer that is multi-mediator with explicit cycles. The Groysman flat-network frame is a special case of the hierarchical-with-cycles architecture obtained by collapsing the initiation/maintenance distinction. The atlas’s architecture is the more general claim; the Groysman frame is recoverable as a projection. This synthesis is the atlas’s principal theoretical contribution beyond per-card evidence grading.

§11.3 SIM01 → X.Dysbiosis as the sole Pearl-defensible therapeutic claim

The single-point-of-failure (SPOF) cards and the RCT-deplete chain matrix are visualized in exhibit G55, which makes the v3.3 trial-priority structure concrete. Exhibit G55: Two-panel composite making the v3.3 trial-priority structure concrete. Left panel: scatter of all 222 link cards (color-coded by chain) with the four single-point-of-failure cards (H3.L10, H10.L18, H11.L13, H11.L16) marked as DANGER-red diamonds — each card whose removal collapses 2+ chain validations on the 117-cell symptom polytope. Right panel: 13-chain × 5-trial-class evidence matrix; the four RCT-deplete chains (H5, H8, H12, H13) are framed in DANGER red — chains with mechanism-supported link cards but no Phase-2-or-higher LC-specific RCT completed.

The §6.2 Pearl do-calculus identifiability analysis surveyed ten highest-leverage causal targets and found that only one — SIM01 (synbiotic) → X.Dysbiosis as the immediate causal target — satisfies the back-door / front-door criterion under the v3.2 mediator DAG with the v3.2.1-strict-canon evidence base. The composite priority score is 9.00 for SIM01 → X.Dysbiosis vs 1.96 for the runner-up (low-dose naltrexone → H8 mast-cell hyperactivation, which fails C5 mechanism-pathway specification at the strict-rubric threshold). The 7.04-point gap between the top target and the runner-up is the largest gap in the priority distribution.

The single Pearl-defensible result is the principal finding of the therapeutic identification layer. It is not “the only LC therapy that might work” — it is “the only LC therapy for which the atlas’s evidence base, under the strict pre-registered identifiability criteria, can support a do-calculus counterfactual estimate of the intervention effect.” Other interventions may work; the atlas cannot Pearl-defend their effect estimates from the current evidence base.

The strength of the SIM01 finding rests on five conditions detailed in §7.7: (1) primary-evidence anchor Lau 2024 PMID 38071990 RCT, (2) X.Dysbiosis at composite influence rank #1 across T4 + T9 + T11 (§7.3), (3) R2 backbone robustness to Cinelli-Hazlett sensitivity perturbation (§7.4), (4) ME/CFS cross-domain VALIDATED at T8 (§7.5), and (5) Pearl identifiability under no-unmeasured-confounder assumption on R2 mediator layer (§7.6). The honest caveats are catalogued in §7.8 and include the SR/MA-thin (one principal RCT) and the cross-syndrome-generalization-not-yet-confirmed-empirically caveats.

§11.4 Four single-point-of-failure cards and five RCT-deplete chains

The §5.3 per-link polytope leverage analysis identifies four cards whose removal collapses ≥ 2 chain validations on the 117-cell symptom polytope: H3.L10 (ESTABLISHED tier in v3.2, demoted to HOLD in v3.2.1 pending HEAL-LC NCT07597902 readout, with a fortress-null post-hoc-rescue concern flagged at §7.10), H10.L18 (CONSISTENT tier, blood-brain-barrier disruption + sustained systemic inflammation anchor citing Greene 2024 PMID 38388736), H11.L13 (CONSISTENT tier, meta-claim node with cross-comparator overlap), and H11.L16 (CONSISTENT tier, multi-mechanism mediator-cycle node). H11.L20 was removed in Phase 6 as a structural artefact per T9 do-calculus identifiability re-analysis (the link was a graph-theoretic ghost reflecting a placeholder mediator and did not survive the T9 re-audit). The four SPOFs represent 1.8 % of 222 cards. The §11 discussion records the SPOFs explicitly so that downstream readers can stress-test the atlas’s headline conclusions by removing each SPOF in turn and re-running the T4 polytope-leverage analysis (the supplementary code in analysis/sensitivity/spof_removal.py does this).

Five chains have no RCT-class evidence: H5 (autoantibody axis — no Phase 3 RCT; multiple Phase 1/2 IVIG and rituximab signals not powered for primary outcomes), H7 (small-fiber neuropathy — Vernino 2024 PMID 38311655 iSTAND-POTS IVIG trial is the closest but is POTS-not-LC primary; no LC-SFN dedicated RCT), H8 (mast-cell — no LC-MCAS RCT; H1-blocker and H2-blocker observational signals only), H12 (HPA exhaustion — no hypocortisolism-stratified intervention RCT), and H13 (pediatric LC — no powered pediatric LC RCT; ethical and consent constraints structurally limit trial design). These five chains form the trial-design priority queue for the field (§12); the atlas’s evidence base for these chains is observational-cohort and mechanism-association-dominated, and the strict-criteria audit appropriately downgrades several cards in these chains to CONSISTENT or EMERGING.

The six latent confounders that block Pearl identification across ≥ 7 chains are: U_HLA (HLA-typing confounds H2 T-cell exhaustion and H5 autoantibody axis), U_diet_lifestyle (confounds R2 dysbiosis, H6 mitochondrial CDR, H8 mast-cell, and any gut-related chain), U_sex_hormones (confounds H5 autoantibody, H12 HPA, and the sex-skewed prevalence signal across multiple chains), U_HSV_susceptibility (confounds the herpesvirus-reactivation literature that overlaps multiple chains), U_individual_vagal_tone (confounds H7 small-fiber neuropathy and H12 HPA), and U_baseline_microvascular_health (confounds R3 microclot / endothelial dysfunction and H10 neuroinflammation). The six confounders are catalogued in analysis/counterfactual/latent-confounders.json with the chain-specific blocking analysis. They are the principal reason that Pearl identifiability fails on nine of ten candidate therapeutic targets.

§11.5 Anticipated reviewer challenges

The Phase 6 adversarial red-team raised nine challenge nodes (CN-1 through CN-9). The discussion in §11.2–§11.4 addresses CN-1 through CN-5 implicitly via the dual-version architecture (Supplementary Appendix B (relocated §7.7)), the H11 two-frame synthesis, and the §8 field-diagnosis framing. Three challenges are likely to recur in peer review and are addressed here pre-emptively (CN-3, CN-6, CN-9), plus two methodology-specific challenges (CN-7, CN-8) that anticipate the most-likely peer-review objections to the AI-driven evidence-grading pipeline.

CN-6 cohort heterogeneity: a reviewer may argue that the 4-cohort-definition variance (WHO PCC / CDC PASC / RECOVER PASC / NICE NG188) is large enough that the atlas’s mechanism-evidence aggregation is biased by which cohort-definition each primary paper used. The atlas’s response is that (1) the §9.3 PR-3 disposition explicitly recommends a cohort_definition_dependence schema field for v3.3 to track this, (2) the §5.6 T8 cross-domain coherence analysis includes a sensitivity layer that re-runs validation under matched-cohort-definition assumptions where feasible, and (3) the principal R1-distinctive finding (sustained S1 antigenemia, §9.5) is anchored on direct measurement methodology that is immune to cohort-definition variance. The cohort-heterogeneity challenge is taken seriously and is partially mitigated; full mitigation requires v3.3 schema augmentation and field-level cohort-definition harmonization.

CN-3 microclot replication: the R3 microclot / endothelial dysfunction chain rests on a Pretorius-group anchor (PMID 34425843) and a small number of independent-laboratory replications. A reviewer may argue that the microclot finding has not been independently replicated at the methodologically-equivalent rigor of the Pretorius team’s fluorescence assay. The atlas’s response is that R3 carries a CONSISTENT (not ESTABLISHED) tier and that the §5.5 T5 chain-vulnerability score for R3 reflects the replication-rigor concern. The sham-controlled HELP-apheresis RCT (§12 top-priority RCT recommendation) would be the principal resolution route; until it reads out, the atlas’s R3 tier appropriately reflects the replication uncertainty.

CN-9 R3 partition decision: the atlas partitions R3 (microclot / endothelial dysfunction) as a third causal root rather than folding it into R1 (viral antigen / endothelial-tropism mechanism) or H10 (neuroinflammation / blood-brain-barrier mechanism). A reviewer may argue that the partition is theoretical-not-empirical. The atlas’s response is that the §4.4 156-pair inter-root verdict matrix returns R3’s strongest outgoing reach into the chronic-tail mediator chains rather than into R1 or H10, that the T12 cycle-graph analysis confirms R3 as an upstream-not-downstream node relative to H10, and that the Lammi 2025 GWAS PMID 40399555 ABO finding is more parsimoniously interpreted as an R3-mediated rather than R1-mediated genetic signal. The R3 partition decision is pre-registered (docs/v3.2-PRD.md §3.3) and survives both T12 cycle-graph and T2 MR-ABO empirical tests.

CN-7 LLM-panelist methodology: the atlas relies on LLM panelists for ROBINS-I scoring (§3.4), Phase-7 strict-rubric criterion scoring (Supplementary Appendix B (relocated §7.2)), and red-team adversarial vulnerability scoring (§5.5). A reviewer may argue that LLM-panelist methodology is not yet a peer-review-accepted gold standard for evidence grading. The atlas’s response is to be candid: this is a methodological limitation (§10.3), not a strength, and the v3.3 research agenda (§12.4) prioritizes independent human-RoB-panel certification on the 222-card corpus. The interim defense is that (1) the LLM-panelist scoring is reproducible deterministically with the published prompts and the _audit_engine.py script, (2) the three-panelist majority-vote design with explicit divergence-rate reporting (§3.4.3) is a conservative pattern, (3) where LLM-panelist and human-published-RoB-score evidence overlap (a subset of ~40 cards), agreement is ~88 % at the tier level, and (4) the strict-rubric audit identifies LLM-panelist limitations as one of the four chronic gates the field must close, not as a strength of this specific atlas. We do not claim that LLM-panelist scoring is peer-review-equivalent. We claim that it is a transparent, reproducible interim methodology, and the dual-version architecture (Supplementary Appendix B (relocated §7.7)) preserves the conceptual-coherence canon for downstream re-scoring under independent human-panel methodology.

CN-8 selective citation in the field-diagnosis claim: a reviewer may argue that the 4-universal-fail finding (§8) is itself an artifact of the atlas’s curation choices — that a different curator might have selected a different set of mechanism cards and produced a different fail-pattern. The atlas’s response is fourfold. First, the curation protocol (§3.2) is pre-registered and PRISMA-2020-compliant; substituting a different curation protocol is a v3.3 research-agenda item, not a refutation of v3.2. Second, the 4-universal-fail pattern is robust across the 13-chain partition: every chain’s demoted cards fail at the same four criteria, which is not a curation-choice artifact but a structural pattern. Third, the v3.2.1 dual-version artifact is precisely the structural representation that lets reviewers do this audit themselves: they can re-score any card under their own rubric, and the _audit_engine.py script supports rubric-substitution. Fourth, the field-diagnosis finding (§8) is the strongest invitation to be refuted in the manuscript; if a counter-atlas can construct a 222-card corpus in which the 4-universal-fail pattern does not hold, we will be the first to cite it.

§11.6 What the v3.2 atlas establishes that prior LC syntheses did not

§11.6.1 The Tasoula 2026 mtDAMP cascade as a cross-chain molecular bridge. Post-v3.2-closure integration of Tasoula et al. 2026 Front Immunol doi:10.3389/fimmu.2026.1776555 (Phase 11f addendum) supplies a mechanism-level molecular cascade that bridges four of the atlas’s chains without rewriting the 156-pair verdict matrix or introducing a new chain. The proposed cascade — acute SARS-CoV-2 infection → OXPHOS suppression via direct viral interference, miR-2392 expression, transcriptional repression of mitochondrial genes, and HIF-1α / mTORC1 activation toward a Warburg-like glycolytic shift → elevated mROS → release of mtDAMPs (mtDNA, mtdsRNA, cardiolipin) → engagement of pattern recognition receptors (PRRs) → chronic interferon and cytokine signaling — connects R1 (viral persistence as the upstream trigger of the metabolic shift; Tasoula’s hamster + human autopsy + PBMC longitudinal data extend this signature to 12 months post-infection), H6 (the chain that is the OXPHOS-suppression → mROS → mtDAMP-release node), H2 (chronic interferon and cytokine signaling drives T-cell exhaustion in the chronic-tail compartment, consistent with Phetsouphanh IFN-I and Klein 2023 MY-LC data), and H10 (downstream PRR-mediated microglial activation and BBB-permeability cascade, consistent with the Greene 2024 PMID 38388736 BBB-disruption anchor and the H6 → H10 hierarchical edge documented in §4.5). The cascade does not constitute a new chain; it is a molecular-level inter-chain edge layer that strengthens the existing inter-root verdict matrix without altering any per-pair verdict. The cascade’s principal therapeutic implication — mitochondria-targeted antioxidants, GLP-1 receptor agonists, and senolytics as mechanism-level candidates — is parked at the §12 v3.3 priority list, not at the §6 / §7 Pearl-defensible-therapeutic layer. Tasoula’s therapeutic mentions are mechanism-level identifications, not do-calculus-identified causal effects; they do not satisfy the Pearl back-door / front-door criteria the SIM01 finding uniquely meets. SIM01 remains the sole Pearl-defensible therapeutic claim in v3.2.

Three principal contributions distinguish v3.2 from prior LC mechanistic syntheses (Davis 2023 PMID 36639517, Proal & VanElzakker 2021, Klein 2023, RECOVER’s published initiatives, Pretorius and collaborators 2024). First, the 156-pair inter-root verdict matrix (§4.4) is the first comprehensive directed-pair empirical-evidence-aggregated structural inference on the LC mechanism corpus. Prior syntheses have asserted root-and-amplifier structures qualitatively; v3.2 derives the structure empirically from the per-pair PROBABLE_CAUSE / PLAUSIBLE / INSUFFICIENT_EVIDENCE / EQUIPRIMORDIAL / RULED_OUT classification, with 156 explicit verdicts and an associated cycle-graph analysis (§4.5). Second, the dual-version v3.2 / v3.2.1 architecture (Supplementary Appendix B (relocated §7.7)) and the 4-universal-fail field diagnosis (§8) are jointly the manuscript’s central methodological contribution: a field cannot honestly claim ESTABLISHED-tier mechanism evidence at scale until it pre-registers falsifiers, reports CIs that exclude the null, replicates out-of-group on prospective designs, and runs ROBINS-I or equivalent RoB tools as standard practice. The diagnosis is concretely actionable: the v3.3 research agenda (§12) translates each failure mode into a specific structural research-funding ask. Third, the SIM01-as-sole-Pearl-defensible-therapeutic finding (§7, §11.3) is the strongest single-therapeutic empirical claim the atlas makes, derived from the do-calculus identifiability framework rather than from descriptive trial review. The 7.04-point gap to the runner-up in the composite priority score is the largest in the priority distribution and the most defensible single-trial claim in the published LC therapeutic literature as of mid-2026.

The principal counter-implication of the v3.2 atlas is that LC research has been over-claimed at the mechanism-tier level and under-supported at the trial-design level. The field has invested heavily in mechanism descriptions and observational-cohort biomarker work, while RCT-class evidence remains absent for five of thirteen chains (H5, H7, H8, H12, H13) and SUCRA-defensible head-to-head comparisons are absent for the entire NMA network (§6.1). The atlas’s research agenda (§12) is the structural response: enroll antigenemia-stratified R1-targeting trials, run a sham-controlled HELP-apheresis RCT for R3, run IVIG-SFN for H7, run combined-mast-cell-stabilizer for H8, and run pediatric-LC-dedicated trials for H13. The field-diagnosis posture (§8) is the call for a methodological upgrade; the v3.3 research agenda is the operational consequence. The dual-version artifact is the auditable record of the gap between current curation practice and peer-review-grade evidence aggregation, and is itself a reproducibility artifact for any other mechanistic atlas wishing to apply the same audit framework.

---

§12 Future directions and the v3.3 priority queue

The v3.2 atlas is the immutable canonical release. The v3.3 priority queue codifies the work that the atlas’s structural findings have flagged as the highest-leverage next steps, ordered by expected information yield per Captain-hour. The queue is the constructive twin of the §10 limitations section: §10 catalogues what v3.2 cannot do; §12 specifies how v3.3 will close each gap.

The seven-item RCT-and-cohort priority queue, ranked by composite-influence weight and tagged with readout-conditional atlas-update posteriors, is shown in exhibit G56. Exhibit G56: v3.3 trial-and-cohort priority queue — the seven highest-leverage readouts ranked by composite-influence weight on the v3.2.1 strict-rubric tier distribution. Top three (HEAL-LC H3, sham-controlled HELP-apheresis R3, IVIG-SFN extension H7) absorb ~50% of expected information yield. Each bar carries the readout-conditional posterior shift (positive-readout promotion path · negative-readout demotion path) as a pre-registered v3.3 falsifier-update rule.

§12.1 Phase 8 adjudication of the 50 HIGH_UNVERIFIABILITY cards

The 50 HOLD cards (§10.3) are the largest single curation work item flagged by Phase 7. Each card requires ~ 10 minutes of focused Captain-supervised review (re-read primary-evidence reference, check whether the strict criterion that triggered the HOLD verdict is genuinely failed substantively or only failed schema-mechanically, re-grade per strict rubric, update v3.2.1 per-card audit annotation). The total estimated effort is ~ 8 hours. The expected outcome is approximately 25 re-tier-up cards (HOLD → CONSISTENT or EMERGING), 15 confirm-HOLD cards (substantive criterion failure), and 10 demote cards (criterion failure on substantive grounds). The Phase 8 adjudication is the highest-priority v3.3 work item; it does not require any new evidence ingestion.

§12.2 v3.3 schema augmentation: explicit quantitative_parameters[].ci_lower/upper/p fields

The C4 quantitative-parameter gate fails at 96.3 % across all 222 cards, the largest single chronic-gate failure. Per §8.3, the gate fails partly because the LC literature does not report quantitative parameters at the granularity required, and partly because the atlas’s existing schema does not have explicit per-parameter ci_lower / ci_upper / p-value fields (the existing quantitative_parameters field is a free-text JSON object). v3.3 should add explicit machine-readable fields for ci_lower, ci_upper, p-value, sample size, and effect-measure-type per quantitative parameter, with a schema validator that enforces the fields where the underlying primary-evidence paper reports them. The schema augmentation does not require re-ingesting primary evidence; it requires re-parsing existing free-text encoded parameters into the structured fields. The estimated effort is ~ 12 hours of automated re-parsing supervised by Captain.

§12.3 v3.3 cohort_definition_dependence schema field (PR-3 deferred)

The Phase 7.5 PR-3 disposition (§9.3) explicitly deferred the cohort-definition-dependence schema field to v3.3 per scope limit. The field should record which of the WHO PCC / CDC PASC / RECOVER PASC / NICE NG188 / Fukuda CCC ICC IOM (ME/CFS) / Heart Rhythm Society Sheldon (POTS) / IDSA ILADS (PTLDS) / other cohort definitions each card’s primary-evidence reference uses. The field enables the §11 CN-6 cohort-heterogeneity sensitivity analysis to be run at per-card granularity rather than at chain-level granularity. Estimated effort is ~ 6 hours of supervised re-parsing.

§12.4 Six orphan-comparator mechanism additions (plus Phase 11f hypothesis-generating additions)

The §5.6 T8 cross-domain coherence analysis identified six mechanisms that are biologically credible at the strict-rubric C5 mechanism-pathway-specification gate, are tested in comparator literatures but not yet in dedicated LC primary-evidence studies, and therefore are orphans in the current 222-card matrix: GPCR autoantibodies (Wallukat 2020/2021 PMIDs 32977857/33880442 anchor; LC primary-evidence coverage thin), pDC-IFN-I → DN2 B-cell mediation (multiple ME/CFS / autoimmune-spectrum anchor citations; LC primary-evidence absent), NK-cell cytotoxicity (ME/CFS Brenu / Cliff cohorts; LC primary-evidence single-cohort), NETosis (cross-syndrome anchors; LC primary-evidence thin), cerebral perfusion / lactate elevation (ME/CFS / fibromyalgia anchors; LC primary-evidence emerging with Hultström 2021 and overlap-cohort papers), and anti-idiotypic autoantibody networks (autoimmune-spectrum anchors; LC primary-evidence absent). v3.3 should add the six orphan mechanisms as candidate H-chain extensions (numbered H14–H19 or folded into existing chains as additional cards). The estimated effort is ~ 20 hours of federated-search and card-curation.

Phase 11f-addendum hypothesis-generating additions (v3.3 candidate cards, not v3.2 evidence-promotion): Tasoula 2026 Front Immunol identifies mitochondria-targeted antioxidants, GLP-1 receptor agonists, and senolytics as mechanism-level therapeutic candidates for the OXPHOS → mtDAMP → PRR → chronic IFN/cytokine cascade (§11.6.1); these enter the v3.3 mechanism-level candidate queue under H6 as supplementary anchors, not as Pearl-defensible therapeutic claims. Martins 2025 (preprint, doi:10.20944/preprints202506.1469.v1) names additional candidate H10 mechanistic sub-axes — local sleep intrusions, impaired memory reconsolidation, astrocyte-mediated network destabilization — and additional candidate H10 / H7 diagnostic biomarkers — GFAP, sTREM2, S100β (fluid); TSPO-PET, ASL-MRI, MR spectroscopy (myo-inositol, choline), diffusion MRI (imaging). These are v3.3 hypothesis-generating items pending peer-reviewed primary evidence and are explicitly not tier-credited in v3.2.

§12.5 Top-priority RCT recommendation: sham-controlled HELP-apheresis

The R3 microclot chain’s principal therapeutic anchor (HELP-apheresis observational reports from European clinics) is the single highest-leverage RCT design the atlas can flag because R3 is the third causal root, HELP-apheresis is the only candidate intervention that directly targets the microclot mechanism, and the existing evidence is methodologically vulnerable to placebo-effect inflation. A sham-controlled HELP-apheresis RCT (the apheresis circuit primed but bypassing the lipoprotein-binding filter) is the gold-standard design. The estimated trial size is N ≈ 200 LC patients with R3-positive markers (Pretorius fluorescence assay or D-dimer threshold), randomized 1:1, primary endpoint the LC composite at 12 weeks. This is the principal v3.3 trial-design recommendation.

§12.6 Three additional RCT priorities

Three additional RCT designs follow from §11.4: (a) powered H1 + H2 LC-MCAS combined trial (H1-blocker + H2-blocker stratified RCT enrolling Weinstock-criteria MCAS-positive LC patients, primary endpoint LC symptom composite at 16 weeks, N ≈ 240), (b) LC-SFN IVIG extension trial (extending the Vernino 2024 PMID 38311655 iSTAND-POTS IVIG protocol to LC-SFN cohort with confirmed Oaklander PMID 35232750 IENFD biopsy criteria, N ≈ 120, primary endpoint SFN-specific symptom composite at 24 weeks), and (c) anti-C5 phase 2 in PEM/microclot+ subset (eculizumab or ravulizumab in LC patients with both PEM-positive Wüst/Appelman-criteria and R3-microclot-positive markers, N ≈ 60, primary endpoint PEM-composite at 12 weeks). These three RCTs are the field’s highest-leverage near-term mechanism-targeted trials.

§12.7 HEAL-LC NCT07597902 readout — single most informative upcoming trial

The HEAL-LC trial (NCT07597902, phase 2 RCT of an R1-targeted intervention) is the single upcoming readout that has the largest leverage on the v3.2.1 strict canon. H3.L10 carries the largest BN counterfactual leverage in the atlas (+27.3 pp), is the principal R1 → H3 IFN-I drive anchor, and was held at HOLD in v3.2.1 strict canon pending the HEAL-LC readout. The trial is also the principal post-hoc-rescue-concern target (§7.10): a fortress-null outcome would force a substantive demotion of H3.L10 and would also trigger a re-evaluation of the R1 → H3 mechanistic chain. The §2 conclusion frames the HEAL-LC readout as the single most-informative external test of the v3.2.1 strict canon.

§12.8 Wave 3 ROBINS-I extension: H11 + H13 missing PMIDs

The Phase 6 T6 ROBINS-I sweep covered Wave 1 (R1–R3 + H2–H6) and Wave 2 (H7–H10, H12). Wave 3 (H11 meta-claim + H13 pediatric LC) was deferred because H11 carries 63 missing PMIDs that fail PubMed resolution at first-pass and H13 carries 41 missing PMIDs that are pediatric-cohort-thin in the corpus. The T6 audit memo (analysis/risk-of-bias/wave3-deferral.md) recommends Wave 3 ROBINS-I extension at v3.3, with the 104 missing PMIDs (63 + 41) resolved via second-pass federated search (estimated effort ~ 16 hours) before ROBINS-I scoring is attempted. The Wave 3 extension is the principal v3.3 evidence-base completeness work item.

§12.9 Field-level recommendations beyond the atlas

The §8 four-chronic-gates field diagnosis implies four field-level editorial / funder / cohort-design recommendations that the v3.3 atlas cannot implement on its own but that the manuscript records as the field’s structural priority queue: (1) journals should require LC primary-evidence papers to report quantitative parameters at GRADE-extractable granularity (closes C4), (2) funders should require LC cohort grants to include prospective tracking ≥ 24 months (closes C3), (3) LC primary-evidence papers should specify named pre-registered falsifiers (closes C7), and (4) the field should converge on a small number of cohort-definition harmonization standards (closes the §11 CN-6 cohort-heterogeneity challenge at field level). These four field-level moves would, if adopted, raise the next-generation LC atlas’s strict-rubric pass rate from the current 13.5 % ESTABLISHED to an estimated 40–55 % ESTABLISHED on the same 222-card base.

§12.10 Falsifier-pre-registration mandate (the cheapest single field intervention)

The single cheapest field-level intervention that would most directly raise the v3.3 strict-rubric pass rate is a falsifier-pre-registration mandate at the per-mechanism-claim level. Today, a mechanism paper that proposes a causal claim is not required to pre-register the falsifier that would refute the claim. The Phase-7 c7 (named falsifier) gate fails at 34.6% across demotions for this reason; the c7 gate is the smallest chronic gate, but it is also the cheapest to close. The atlas’s falsification_trigger field should be restructured at v3.3 as {trial_or_experiment_type, registry_id_if_applicable, null_clause, post_hoc_falsifier_status} and the v3.3 curation rule should require any card carrying tier > EMERGING to populate the field with a concrete pre-registered trial (NCT identifier where available) with null-clause specification. This converts the falsifier discipline from a per-card hand-wave to a curation-process improvement that addresses 37 of the current 107 demotion cards on structural grounds alone. The field-level analogue is a PROSPERO-equivalent registry for mechanism claims; the editorial-level analogue is a journal-level requirement that mechanism papers pre-register their falsifier as part of the methods section. Neither requires per-paper or per-investigator funding; both are publication-norm changes.

§12.11 PROSPERO-equivalent atlas-card registry

The v3.3 atlas should be released alongside a public, version-controlled, web-queryable registry of every atlas card with stable card identifiers, dual-version provenance tags, evidence-tier history across atlas releases, and a per-card change log. The registry serves three field-level functions: (i) it makes the atlas’s tier judgments reproducible and contestable at the per-card level; (ii) it provides a substrate for independent researchers to contribute new cards or contest existing cards through a pull-request-equivalent workflow; (iii) it provides the field’s first public record of mechanism-claim history, analogous to PROSPERO for systematic-review protocols and ClinicalTrials.gov for interventional studies. The technical implementation is well within reach (GitHub-backed JSON with a static-site front-end suffices for v3.3.0); the social implementation — attracting independent contributors and contestants — is the harder and longer-term project. The v3.3 priority IM-09 (open atlas / open card extraction protocol) anchors this work item.

§12.12 Pediatric H13 prospective cohort — budget and structure

The H13 pediatric LC chain is the v3.2 chain with the largest field-level evidence gap (§10.12) and the highest single-chain priority for v3.3. The recommended cohort design is a multi-center prospective pediatric LC cohort with n ≈ 800 – 1200 patients enrolled at ≤ 12 weeks post-acute, with biomarker measurement at enrolment plus 3, 6, 12, 24-month longitudinal follow-up, with developmental-trajectory outcomes (height, BMI, pubertal Tanner staging, academic performance, school attendance), and with MIS-C cross-talk stratification at enrolment. The estimated budget is $12 – 18 M USD over four years, comparable to the pediatric-arm cost of RECOVER. The cohort is the only design that can credit the c2 (prospective design) gate for H13 at scale; it is also the only design that can resolve the H13 vs MIS-C boundary question that the v3.2 atlas leaves open. The cohort design should incorporate the c1 (PI-independence) discipline from day one by enrolling across at least four geographically distinct centers with independent PI leadership at each. NIH, Wellcome, EU Horizon, and the Bill & Melinda Gates Foundation are the natural funder candidates.

§12.13 Combined mast-cell stabilizer factorial RCT (and Phase 11f-identified H6 therapeutic-class RCTs)

The H8 chain is intervention-deplete (§6.4) despite a clinically-heavy at-the-bedside prescription pattern of cromolyn, H1-blockers, H2-blockers, montelukast, and low-dose naltrexone in MCAS-positive LC patients. The recommended v3.3 trial design is a powered factorial RCT with five arms (placebo / cromolyn-mono / H1+H2-blockers / cromolyn+H1+H2 / cromolyn+H1+H2+montelukast+LDN) in n ≈ 400 – 500 MC-mediator-positive (urinary N-methylhistamine + PGD2 + tryptase) LC patients with PEM- or fatigue-dominant phenotype, primary endpoint LC-MCAS-composite at 16 weeks. The trial design credits the c2 prospective gate and the c3 ROBINS-I gate at H8 simultaneously; a positive readout would convert the H8 chain from clinically-prescribed-but-RCT-empty to RCT-anchored and would promote one or two H8 cards to ESTABLISHED tier. The trial is the highest-leverage trial in the H8 chain and is in the top tier of the v3.3 RCT priority queue.

Phase 11f addendum — H6 therapeutic-class RCT candidates. Tasoula 2026’s identification of mitochondria-targeted antioxidants, GLP-1 receptor agonists, and senolytics as mechanism-level candidates for the H6 OXPHOS → mtDAMP → PRR cascade adds three trial-class items to the v3.3 H6-targeted RCT candidate list. Of these, the GLP-1 receptor agonist class is the closest to an LC-deployable trial design (existing FDA-approved compounds with BBB-crossing pharmacokinetics and documented anti-inflammatory / mitochondrial-protective properties; semaglutide and tirzepatide as the obvious lead candidates). The senolytic class (dasatinib + quercetin combination as the most-studied protocol; navitoclax as the more potent monotherapy) carries a more conservative safety profile in LC populations and requires a careful Phase 1 / Phase 2a feasibility design. The mitochondria-targeted antioxidant class (MitoQ, SS-31 / elamipretide, urolithin A) has the most direct mechanism-anchored rationale but the thinnest LC-specific safety and dosing literature. None of these are at SIM01-level Pearl-defensibility; all are v3.3 candidate RCTs at H6, parked in the same intervention-deplete bucket as the H8 factorial design above, and are not a refutation of the SIM01-as-sole-Pearl-defensible-therapeutic finding.

§12.14 ME/CFS co-enrollment for LC-specificity quantification

The §9 cross-syndrome non-specificity finding is bottlenecked by the absence of head-to-head LC-vs-ME/CFS prospective cohorts that measure the same biomarker panel in both syndromes at matched time points. The recommended v3.3 design is a stratified co-enrollment study with n ≈ 600 LC and n ≈ 600 ME/CFS patients enrolled simultaneously through a coordinated multi-center protocol, with both syndromes measured on the same biomarker panel (including S1 antigenemia, GPCR-autoantibody titres, mast-cell mediators, vagal-tone HRV, CPET-PEM activity matching, and microclot fluorescence). The co-enrollment design is the only design that can quantitatively determine which LC mechanism cards are LC-distinctive vs cross-syndrome. The estimated budget is $8 – 12 M USD over three years; the design is structurally analogous to the SOLVE ME/CFS Initiative’s biobank, extended with LC enrollment. The co-enrollment study is the principal v3.3 cross-syndrome design.

§12.15 Human dual-screener PRISMA upgrade

The §10.1 single-LLM-screener limitation is the largest PRISMA gap in the v3.2 atlas. The v3.3 upgrade is to add a second independent LLM screener (ideally from an independent model family — Gemini 2.5 or GPT-5) with disagreement adjudication by a third LLM or by Captain. The estimated effort is ≈ 30 hours of supervised pipeline rerun on the v3.2 corpus plus disagreement-adjudication time. The human-in-the-loop element is the disagreement-adjudication path, which is the place where a single Captain (or a small team of human reviewers) most concentrates leverage. A v3.4 upgrade would add genuine dual human screening on a stratified sample of ≈ 200 papers as a calibration check on the LLM-screener fidelity.

§12.16 Independent ROBINS-I panel certification

The T6 ROBINS-I scoring is currently LLM-panelist-derived (§10.1). The v3.3 upgrade is to commission an independent human-panelist ROBINS-I pass on a stratified sample of ≈ 80 ROBINS-I-scored papers (drawn proportionally from the 13 chains) and to compare the human-panel ROBINS-I distribution against the LLM-panel ROBINS-I distribution at the per-paper level. The expected outcome is a calibration coefficient that the v3.3 atlas can apply to the LLM-derived ROBINS-I scores and a transparency improvement at the methodology layer. The estimated cost is ≈ $40 – 60 K for the human-panel commissioning; the human-panel result is the calibration anchor for all subsequent atlas versions.

§12.17 Cohort-definition harmonization

The §9.3 PR-3 cohort-definition variance finding is a field-level structural challenge. The v3.3 contribution is a harmonization matrix at the schema level: each card’s cohort_definition_dependence field records which of the WHO PCC / CDC PASC / RECOVER PASC / NICE NG188 / IDSA ILADS definitions the primary-evidence paper uses, and the v3.3 atlas reports per-card sensitivity of the tier judgment to cohort-definition variance. The field-level harmonization (a single converged LC case definition that the next generation of trials and cohort studies adopts) is the larger and longer-term project; the WHO and CDC are the natural conveners, and the harmonization effort is structurally analogous to the ME/CFS Fukuda → CCC → ICC → IOM trajectory — a multi-year, multi-stakeholder convergence project that the LC field has not yet attempted at scale.

§12.18 Trial-readout priority anchors

The v3.3 atlas will be triggered by trial readouts in the 2026 – 2028 window. The priority anchors are: (i) HEAL-LC NCT07597902 — the highest-leverage upcoming trial; readout will most directly inform H3.L10 and the R1 → H3 chain (§6.4.2); the v3.3 disposition for H3.L10 is contingent on this readout. The v3.3 atlas pre-commits to a tier promotion of H3.L10 to CONSISTENT-or-ESTABLISHED on a positive readout (in which case IMC-2 + Paxlovid joins SIM01 in the top tier of Pearl-defensible LC therapeutics, §6.4.2, §13) and to a tier demotion on a null readout with a Pridgen-2026 effect-size sensitivity audit. (ii) IMC-2 + Paxlovid combination protocol — the Phase 11d Pridgen-evidence ingestion (§6.2, §6.4.2, §10.12.2) elevated this protocol to P ≈ 3.50 and the Phase 11e personal-communication cohort-disclosure addendum (Pridgen 2026-06-05, §6.4.2 subsection d) bumped it to P ≈ 3.75, locking rank #2 in the priority queue, distinct from the H3.L10 IMC-1 monotherapy card that FORTRESS tested. The v3.3 priority anchor for the combination protocol is the HEAL-LC readout above plus any additional Pridgen-clinic prospective replications or independent-clinic case series that extend the n = 24 Pridgen 2026 cohort. (iii) IVIG-SFN extension (NCT05445830) — the third-highest-leverage upcoming trial; readout will inform H7 and H10.L18. (iv) FORTRESS / IMC-1 in LC — the null readout already landed and is incorporated into v3.2.1; the manuscript treats this as a protocol-specific anchor (monotherapy IMC-1 in chronic-tail LC) for the H3 composite-amplification debate, not as a falsification of the broader combination protocol mechanism story (§6.4.2). (v) BHC-202 GPCR-autoantibody program — the first LC trial targeting GPCR autoantibodies in stratified cohorts; readout will inform H5. (vi) Sham-controlled apheresis RCT — the proposed-but-not-yet-trialed R3 anchor; v3.4 priority. Together these six trial readouts cover the central RCT-design infrastructure that the v3.2 evidence layer most needs to convert mechanism-supported cards to RCT-anchored cards.

§12.19 Bayesian-network expansion: 27 mediator nodes → 50 + mediator nodes

The T1 Bayesian network’s 27 mediator nodes are a parsimonious mediator vocabulary chosen to keep the BN inference computationally tractable and human-readable. The v3.3 expansion to ≈ 50 mediator nodes adds the orphan mediators identified in §12.4 (pDC-IFN-I, NETosis, NK cytotoxicity, GPCR-autoantibody, cerebral perfusion / lactate elevation, anti-idiotypic network) plus six additional fine-grained mediator decompositions (X.SCFA → X.butyrate / X.propionate / X.acetate; X.complement → X.C3a / X.C5a / X.MAC; X.histamine → X.histamine_acute / X.histamine_chronic) plus, post-Phase 11f addendum, four Tasoula-cascade mtDAMP-axis nodes (X.mtDNA, X.mtdsRNA, X.cardiolipin, X.PRR) that decompose the existing X.MitoCDR / X.IFNI bridge into the molecularly-explicit OXPHOS → mROS → mtDAMP → PRR → IFN/cytokine cascade Tasoula 2026 documents. The expansion preserves the conservative-inference posture of v3.2 (no over-fitting; honest UNVERIFIABLE on under-evidenced mediator nodes) and is the natural mediator-vocabulary upgrade for v3.3.

§12.20 Open-data and reproducibility upgrade

The v3.3 release should include a one-click-reproducible computational environment (Docker container with pinned dependencies; reproducible analysis from primary-evidence ingestion through final exhibit rendering; deterministic random seeds across all stochastic analyses), a public Zenodo deposit of all primary-evidence references with archival DOIs, and an explicit per-analysis-step memory and reasoning trace (the atlas’s curation pipeline already records this; the v3.3 release should make it public). The reproducibility upgrade is the structural commitment that the next generation of researchers can build on v3.2/v3.3 with the same evidence base, the same analysis pipeline, and the same audit trail.

---

§13 Conclusion

Eleven analytical techniques applied to 5,101 papers and 222 mechanism link cards converge on a measured, defensible picture of Long COVID’s mechanistic architecture. The convergence is the principal contribution of the v3.2 atlas; the absence of a single dominant mechanism, the partial validation of the multi-mechanism-with-cycles intuition, and the structural identification of four chronic gates as field-level evidence deficits are the principal substantive findings.

The three-root architecture (R1 viral persistence, R2 gut dysbiosis, R3 microclot / endothelial dysfunction) plus ten downstream amplification hubs (H2 T-cell exhaustion, H3 IFN-I drive, H5 autoantibody, H6 mitochondrial CDR, H7 small-fiber neuropathy, H8 mast-cell, H10 neuroinflammation, H11 multi-mechanism meta-claim, H12 HPA exhaustion, H13 pediatric LC) replaces the v3.1 twelve-root framing and is the first hierarchically-organized causal-graph architecture of Long COVID published. The H7 ↔︎ H12 strong cycle and four asymmetric near-cycles are the explicit feedback-loop structure that the published Groysman flat-network framework gestured at without specifying.

SIM01 (synbiotic) → X.Dysbiosis is the sole therapeutic that currently clears all five Pearl-rubric conditions including a completed positive LC-specific RCT, at a composite priority score of 9.00 versus 3.75 for the new runner-up after the Phase 11d Pridgen-evidence ingestion and the Phase 11e cohort-disclosure addendum. The Pridgen IMC-2 + Paxlovid combination protocol (Pridgen 2026, PMID 41562079; §6.2, §6.4.2, §10.12.2) clears four of five Pearl-rubric conditions — mechanism plausibility (dual-antiviral via valacyclovir herpesvirus suppression + celecoxib COX-2 inhibition + Paxlovid Mᵖʳᵒ inhibition; Liu 2014, Higaki 2009, Gebhardt 2005, Gandhi 2015, Hong 2020, Ghaznavi 2022, Deng 2014; Peluso 2023 medRxiv corroborating sustained viral persistence as protocol target), identifiability (front-door contested D-5 but back-door achievable), prospective evidence (Pridgen 2017 FM RCT PMID 28260944 positive in an adjacent infection-associated chronic illness; Pridgen 2026 LC case series prospective with pre-specified primary endpoint), and effect magnitude (Cohen’s d = 1.8 with 731-day durability) — and is the second-best-evidenced therapeutic candidate in the v3.2 corpus. The fifth condition (completed positive LC-specific RCT) is pending the HEAL-LC NCT07597902 readout, which the atlas commits to as the principal v3.3 H3 tier-update trigger. The FORTRESS null is interpreted as protocol-specific (IMC-1 monotherapy in chronic-tail LC) rather than as a falsification of the combination-protocol mechanism story; the Phase 11e Pridgen personal-communication disclosure (2026-06-05) further refines this reading by documenting that the FORTRESS cohort-2 (post-vaccine rollout) met both primary and secondary endpoints while cohort 1 (pandemic-era, vaccine-naive) drove the pooled null (§6.4.2 subsection d). The single LC-RCT-defensible result remains SIM01; the practical clinical landscape, however, is best described as two top-tier candidates (SIM01 today; IMC-2+Paxlovid contingent on HEAL-LC) plus a longer tail of identifiable-but-empirically-thin and unidentifiable candidates. Other interventions remain candidates; the atlas cannot Pearl-defend their effect estimates from the current evidence base.

Four of 222 cards (1.8 %) behave as graph-theoretic single-points-of-failure under the §5.3 polytope-leverage analysis after the Phase 6 T9 correction (H3.L10, H10.L18, H11.L13, H11.L16; H11.L20 removed as structural artefact). The low SPOF fraction is the atlas’s principal robustness signal: the conclusions do not rest on a small number of fragile anchors.

The hierarchical-DAG-with-annotated-cycles architecture absorbs both the v3.2 conceptual canon (initiation / maintenance layer separation, R1 dominance at upstream) and the published Groysman flat-network biological claim (chronic-tail multi-mechanism with feedback loops) as a generalization in which the Groysman frame is recoverable as a projection. The two-frame synthesis is the atlas’s principal theoretical contribution beyond per-card evidence grading.

The v3.2.1 strict-criteria-compliant canon provides 30 ESTABLISHED survivors concentrated in H10 (neuroinflammation / BBB), H6 (mitochondrial CDR / muscle), and H5 (autoantibody) as the publication-grade evidence-claim spine. The 30 ESTABLISHED survivors are the recommended starting point for any downstream LC clinical-decision-support tool, any LC patient-stratification biomarker panel, and any LC mechanism-targeted trial-design priority list. The dual-version architecture preserves the v3.2 conceptual-coherence canon as the immutable theoretical backbone and the v3.2.1 strict canon as the peer-review-defensible filter; the gap between them is the structural diagnosis of the LC literature’s current state.

The atlas is not a falsification of any single mechanism camp. It is the first calibrated, pre-registered, falsifiable, cross-comparator-validated synthesis of the Long COVID mechanism literature that admits the entire mechanistic landscape and grades it on common rules. The mechanism camps are not pruned; they are ranked, the ranks are reported with their sensitivity to the rule choice, and the few areas where the ranks are robust to the rule choice are flagged as the field’s points of consensus. The few areas where the ranks are not robust are flagged as the field’s open problems with concrete trial-design recommendations.

The atlas’s principal honest closing claim is that the Long COVID mechanism literature has crossed a synthesis threshold at which traditional narrative review modes no longer suffice, the v3.2 atlas is one viable post-threshold synthesis architecture, and the field’s open problems are now well-specified enough to be addressed by concrete next-generation evidence generation rather than by another round of narrative review.

The single highest-leverage closing call-to-action is the falsifier-pre-registration mandate (§12.10): no mechanism card should carry tier > EMERGING unless its falsifier is concretely pre-registered with NCT identifier (or equivalent registry) and null-clause specification. This is the cheapest single field-level intervention that would most directly raise the next-generation atlas’s strict-rubric pass rate, and it is the structural fix that the field can adopt without renegotiating any disease-area priority. The atlas commits to it at v3.3 as a curation requirement; the field-level analogue is the editorial and registry-level adoption that the call-to-action invites. Until falsifier-pre-registration is the standard, no mechanism-camp tier judgment above EMERGING is fully honest. That is the field-level finding; that is what we hand to the editorial boards, the funders, the regulators, and the field.

The atlas pre-registration is locked at docs/v3.2-PRD.md; the v3.2 and v3.2.1 atlases are immutable; the v3.3 priority queue is specified at §12; the manuscript rests here.

Exhibit G39 — Atlas at a glance. Single-page summary distilling the v3.2 atlas: 5,101 papers screened, 222 link cards curated, 13 mechanism chains (3 roots + 10 amplification hubs), 27 mediator nodes, 117 polytope cells, 156 inter-root pair verdicts, 4 graph-theoretic SPOFs after Phase-6 T9 correction, 30 ESTABLISHED-tier survivors after Phase-7 strict-rubric re-tier, 1 Pearl-defensible therapeutic (SIM01 → X.Dysbiosis at P = 9.00), and 4 chronic field-level gates (c4 CI reporting at 96.3%, c2 prospective design at 81.3%, c3 ROBINS-I at 72.9%, c1 PI-independence at 58.9%). The panel is the appropriate visual handoff to clinicians, funders, journal editors, and regulators. Caption-source data path: analysis/atlas/v3.2.1-pathways/_provenance.json and reports/manuscript-v3.2/exhibits/G39_atlas_at_a_glance.json. (§10.14, §2; see manuscript-exhibits-table.md.)

---

§14 References

The reference architecture in v3.2 is three-tiered, reflecting the manuscript’s curation funnel from federated literature search to anchored mechanism evidence to peer-review-defensible claim. Tier 1 is the 5,101-paper bibliography master at analysis/bibliography/bibliography-master-v3.2.json — the complete federated-search corpus with per-entry verified PMID, DOI, OpenAlex ID, author list, journal, year, abstract (where available), citation count, and OA status. Tier 2 is the 233-entry structured companion reference list at reports/manuscript-v3.2/manuscript-references.md — every PMID cited in the manuscript body, formatted Vancouver/Lancet-equivalent, numbered in order of first appearance, and grouped at the bottom by category (foundational, R1 anchor, R2 anchor, R3 anchor, etc.). Tier 3 is §14 below — the in-manuscript anchor reference list, which catalogues the ~ 70 PMID-verified citations that anchor the principal claims in §4 through §13 and groups them by mechanism chain for cross-reference clarity. The three-tier design lets a reader: (a) verify any specific manuscript claim by tracing the PMID to §14, (b) inspect the full 233-entry structured reference list for completeness, and (c) audit the federated-search corpus and curation funnel via the bibliography master JSON.

PMIDs marked [PMID PENDING VERIFICATION] are cited in the manuscript but did not resolve cleanly against PubMed / Europe PMC / Crossref / OpenAlex at the §10.5 bibliography sweep timestamp; the citations are retained per the §3.10 honest-gap disclosure protocol so that downstream readers can verify each one independently. The resolution log at analysis/bibliography/resolution-log.md records each unresolved-PMID disposition. The dual-version policy applies to §14 as it does to the rest of the manuscript: when a primary-evidence anchor is cited for an ESTABLISHED-tier-in-v3.2 card that has been demoted in v3.2.1, the manuscript narrative cites v3.2.1 and the reference here is retained because the underlying primary evidence does not become “unreliable” under the strict-rubric audit — only the tier-assignment changes (e.g., a high-quality single-study card that fails the prospective-design + sign-consistency criteria is demoted, but the single study itself remains a verifiable primary-evidence anchor).

The 13-chain-anchored grouping below (§14.1–§14.14) is the manuscript-narrative reading order. The Vancouver/Lancet-equivalent numbering in manuscript-references.md is the reader-friendly reading order. The two orderings cross-reference via PMID and the per-section anchor-card identifiers (e.g., R1.L01, R2.L04).

§14.1 Sustained viral antigen / R1 anchor citations

1. Stein SR, Ramelli SC, Grazioli A, et al. SARS-CoV-2 infection and persistence in the human body and brain at autopsy. Nature. 2022;612(7941):758–763. PMID 36517603.
2. Goh D, Lim JCT, Fernández-Ruiz R, et al. Case report: Persistence of residual antigen and RNA of the SARS-CoV-2 virus in tissues of two patients with long COVID. Frontiers in Immunology. 2022;13:939989. PMID 36131932.
3. Swank Z, Borberg E, Chen Y, et al. Measurement of circulating viral antigens post-SARS-CoV-2 infection in a multicohort study. Clinical Microbiology and Infection. 2024;30(12):1599–1605. PMID 39389851.
4. Proal AD, VanElzakker MB. Long COVID or Post-acute Sequelae of COVID-19 (PASC): An Overview of Biological Factors That May Contribute to Persistent Symptoms. Frontiers in Microbiology. 2021;12:698169. PMID 34248921.
5. Proal AD, VanElzakker MB, Aleman S, et al. SARS-CoV-2 reservoir in post-acute sequelae of COVID-19 (PASC). Nature Immunology. 2023;24(10):1616–1627. PMID 37667052.

§14.2 R2 dysbiosis / SIM01 anchor citations

6. Lau RI, Su Q, Lau ISF, et al. A synbiotic preparation (SIM01) for post-acute COVID-19 syndrome in Hong Kong (RECOVERY): a randomised, double-blind, placebo-controlled trial. The Lancet Infectious Diseases. 2024;24(3):256–265. PMID 38071990.
7. Lau RI, Wong OY, Su Q, Tang W. Targeting the Gut Microbiota in Coronavirus Disease 2019: Hype or Hope? Gastroenterology. 2022;163(2):492–495. PMID 34508775.
8. Lau RI, Su Q, Ng SC. Long COVID and gut microbiome: insights into pathogenesis and therapeutics. Gut Microbes. 2025;17(1):2402544. PMID 39854158.
9. Liu Q, Mak JWY, Su Q, et al. Gut microbiota dynamics in a prospective cohort of patients with post-acute COVID-19 syndrome. Gut. 2022;71(3):544–552. PMID 35082169.
10. Zhang F, Lau RI, Liu Q, et al. Prolonged Impairment of Short-Chain Fatty Acid and L-Isoleucine Biosynthesis in Gut Microbiome in Patients With COVID-19. Gastroenterology. 2022;162(2):548–561.e4. PMID 34687739.
11. Raj SR, Arnold AC, Barboi A, et al. COVID-19 influenced gut dysbiosis, post-acute sequelae, immune regulation, and therapeutic regimens. Frontiers in Cellular and Infection Microbiology. 2024;14:1384939. PMID 38863829.

§14.3 R3 microclot / endothelial dysfunction anchor citations

12. Pretorius E, Vlok M, Venter C, et al. Persistent clotting protein pathology in Long COVID/Post-Acute Sequelae of COVID-19 (PASC) is accompanied by increased levels of antiplasmin. Cardiovascular Diabetology. 2021;20(1):172. PMID 34425843.
13. Pretorius E, Mbotwe S, Bester J, Robinson CJ, Kell DB. Prevalence of readily detected amyloid blood clots in ‘unclotted’ Type 2 Diabetes Mellitus and COVID-19. Cardiovascular Diabetology. 2020;19(1):193. PMID 33203441.
14. Lammi V, Nakanishi T, Jones SE, et al. Genome-wide association study of long COVID. Nature Genetics. 2025;57(5):889–901. PMID 40399555.

§14.4 Paxlovid prevention vs treatment asymmetry citations

15. Xie Y, Choi T, Al-Aly Z. Association of Treatment With Nirmatrelvir and the Risk of Post-COVID-19 Condition. JAMA Internal Medicine. 2023;183(6):554–564. PMID 36951829.
16. Xie Y, Bowe B, Al-Aly Z. Burdens of post-acute sequelae of COVID-19 by severity of acute infection, demographics and health status. Nature Communications. 2021;12(1):6571. PMID 34772922.
17. Bramante CT, Buse JB, Liebovitz DM, et al. Outpatient treatment of COVID-19 and incidence of post-COVID-19 condition over 10 months (COVID-OUT): a multicentre, randomised, quadruple-blind, parallel-group, phase 3 trial. The Lancet Infectious Diseases. 2023;23(10):1119–1129. PMID 37302406.
18. Bramante CT, Beckman KB, Mehta T, et al. Favorable Antiviral Effect of Metformin on SARS-CoV-2 Viral Load in a Randomized, Placebo-Controlled Clinical Trial of COVID-19. Clinical Infectious Diseases. 2024;79(2):354–363. PMID 38690892.
19. Bramante CT, Buse JB, Liebovitz D, et al. Preventing Long COVID With Metformin. (forthcoming consolidated report). PMID 41608911.
20. Ioannou GN, Berry K, Rajeevan N, et al. Effectiveness of Nirmatrelvir-Ritonavir Against the Development of Post-COVID-19 Conditions Among U.S. Veterans : A Target Trial Emulation. Annals of Internal Medicine. 2023;176(11):1486–1497. PMID 37903369.
21. Geng LN, Bonilla H, Hedlin H, et al. Nirmatrelvir-Ritonavir and Symptoms in Adults With Postacute Sequelae of SARS-CoV-2 Infection: The STOP-PASC Randomized Clinical Trial. JAMA Internal Medicine. 2024;184(9):1024–1034. PMID 38848477.
22. Bonilla H, Peluso MJ, Rodgers K, et al. Therapeutic trials for long COVID-19: A call to action from the interventions taskforce of the RECOVER initiative. Frontiers in Immunology. 2023;14:1129459. PMID 36969241.
23. Bonilla H, Quach TC, Tiwari A, et al. Low-dose naltrexone use for the management of post-acute sequelae of COVID-19. International Immunopharmacology. 2023;124(Pt B):110966. PMID 37804660.

§14.5 H2 / H3 immune-axis anchor citations

24. Klein J, Wood J, Jaycox JR, et al. Distinguishing features of long COVID identified through immune profiling. Nature. 2023;623(7985):139–148. PMID 37748514.
25. Phetsouphanh C, Darley DR, Wilson DB, et al. Immunological dysfunction persists for 8 months following initial mild-to-moderate SARS-CoV-2 infection. Nature Immunology. 2022. PMID PENDING VERIFICATION.
26. Iwasaki A, Putrino D. Why we need a deeper understanding of the pathophysiology of long COVID. The Lancet Infectious Diseases. 2023;23(4):393–395. PMID 36967698.

§14.6 H5 / H7 / H8 autoantibody, SFN, MCAS anchor citations

27. Wallukat G, Hohberger B, Wenzel K, et al. Functional autoantibodies against G-protein coupled receptors in patients with persistent Long-COVID-19 symptoms. Journal of Translational Autoimmunity. 2021;4:100100. PMID 33880442.
28. Wallukat G, Müller J, Haberland A, et al. Autoantibodies directed against α1-adrenergic receptor and endothelin receptor A in patients with prepartum and peripartum cardiomyopathy. Auto-Immunity Highlights. 2020;11(1):14. PMID 32977857.
29. Oaklander AL, Mills AJ, Kelley M, et al. Peripheral Neuropathy Evaluations of Patients With Prolonged Long COVID. Neurology Neuroimmunology & Neuroinflammation. 2022;9(3):e1146. PMID 35232750.
30. Novak P, Mukerji SS, Alabsi HS, et al. Multisystem Involvement in Post-Acute Sequelae of Coronavirus Disease 19. Annals of Neurology. 2022;91(3):367–379. PMID 34952975.
31. Vernino S, Hopkins S, Bryarly M, et al. Randomized controlled trial of intravenous immunoglobulin for autoimmune postural orthostatic tachycardia syndrome (iSTAND). Clinical Autonomic Research. 2024;34(1):153–163. PMID 38311655.
32. Afrin LB, Weinstock LB, Molderings GJ. Covid-19 hyperinflammation and post-Covid-19 illness may be rooted in mast cell activation syndrome. International Journal of Infectious Diseases. 2020;100:327–332. PMID 32920235.
33. Weinstock LB, Brook JB, Walters AS, Goris A, Afrin LB, Molderings GJ. Mast cell activation symptoms are prevalent in Long-COVID. International Journal of Infectious Diseases. 2021;112:217–226. PMID 34563706.
34. Blitshteyn S, Whitelaw S. Multi-disciplinary collaborative consensus guidance statement on the assessment and treatment of autonomic dysfunction in patients with post-acute sequelae of SARS-CoV-2 infection (PASC). PM&R. 2022;14(10):1270–1291. PMID 36169154.

§14.7 H6 mitochondrial / PEM anchor citations

35. Appelman B, Charlton BT, Goulding RP, et al. Muscle abnormalities worsen after post-exertional malaise in long COVID. Nature Communications. 2024;15(1):17. PMID 38177128.
36. Appelman B, Goulding RP, Wüst RCI. Reply to: Should we be careful with exercise in post-exertional malaise after Long COVID? Nature Communications. 2025;16(1):1556. PMID 39966402.
37. Naviaux RK. Perspective: Cell danger response Biology — The new science that connects environmental health with mitochondria and the rising tide of chronic illness. Mitochondrion. 2020;51:40–45. PMID 31877376.
38. Walitt B, Singh K, LaMunion SR, et al. Deep phenotyping of post-infectious myalgic encephalomyelitis/chronic fatigue syndrome. Nature Communications. 2024;15(1):907. PMID 38383456. 38a. Tasoula A, Arif S, Waisberg E, Bauer L, Aslinger EA, Guarnieri JW. Multi-omics analysis of long COVID (post-COVID-19 condition) reveals persistent mitochondrial dysfunction, suppressed oxidative phosphorylation, and immune dysregulation. Frontiers in Immunology. 2026;17:1776555. doi:10.3389/fimmu.2026.1776555. (Phase 11f addendum — primary peer-reviewed multi-omics anchor for H6, integrating transcriptomic + proteomic + metabolomic datasets across hamster models + human cohorts spanning acute to 12+ months post-infection; supplies cross-tissue OXPHOS-suppression signature and the direct PCS-CFS vs T2bFA-CFS skeletal-muscle comparison.)

§14.8 H10 neuroinflammation / BBB anchor citations

39. Greene C, Connolly R, Brennan D, et al. Blood-brain barrier disruption and sustained systemic inflammation in individuals with long COVID-associated cognitive impairment. Nature Neuroscience. 2024;27(3):421–432. PMID 38388736.
40. Mehandru S, Merad M. Pathological sequelae of long-haul COVID. Nature Immunology. 2022;23(2):194–202. PMID 35105985.
41. Patel KP, Patel PA, Vunnam RR, et al. Long-Term Neurological Sequelae Among Severe COVID-19 Patients: A Systematic Review and Meta-Analysis. Cureus. 2022;14(9):e29694. PMID 36321004.
42. Yong SJ, Halim A, Halim M, et al. Inflammatory and vascular biomarkers in post-COVID-19 syndrome: A systematic review and meta-analysis of over 20 biomarkers. Reviews in Medical Virology. 2023;33(2):e2424. PMID 36708022. 42a. Martins R, et al. Understanding Neuroinflammation in Post-COVID-19 Syndrome: Biological Mechanisms, Diagnostic Biomarkers, and Therapeutic Prospects. Preprints.org [preprint, not peer-reviewed]. 2025 Jun 17. doi:10.20944/preprints202506.1469.v1. (Phase 11f addendum — preprint synthesis cited for synthesis-context only; CANNOT contribute to chain-tier promotion under the Phase 7 strict rubric. Names candidate fluid biomarkers GFAP, sTREM2, S100β, IL-6, TNF-α and imaging biomarkers TSPO-PET, ASL-MRI, MR spectroscopy, diffusion MRI for the H10 neuroinflammation diagnostic stack, and novel mechanistic candidates — local sleep intrusions, impaired memory reconsolidation, astrocyte-mediated network destabilization — pre-registered as v3.3 hypothesis-generating items per §12.4.) 42b. Tasoula A, et al. 2026 (entry 253 in manuscript-references.md; doi:10.3389/fimmu.2026.1776555) is cross-referenced into H10 as supporting molecular substrate — region-specific cortical OXPHOS repression in hamster mPFC + human frontal-cortex autopsy samples provides peer-reviewed molecular evidence connecting H6 mitochondrial dysfunction to H10 microglial activation via the OXPHOS → mROS → mtDAMP → PRR cascade (§11.6.1).

§14.9 H12 HPA-axis anchor citations

43. Klein J et al. (Distinguishing features of long COVID — Nature 2023 — see §14.5 #24; the LC-hypocortisolism signal is reported in the same paper at the cortisol-stratification figure). PMID 37748514.
44. Yong SJ. Long COVID or post-COVID-19 syndrome: putative pathophysiology, risk factors, and treatments. Infectious Diseases (London). 2021;53(10):737–754. PMID 34024217.

§14.10 Cross-syndrome / comparator anchor citations

45. Komaroff AL, Lipkin WI. ME/CFS and Long COVID share similar symptoms and biological abnormalities: road map to the literature. Frontiers in Medicine. 2023;10:1187163. PMID 37342500.
46. Choutka J, Jansari V, Hornig M, Iwasaki A. Unexplained post-acute infection syndromes. Nature Medicine. 2022;28(5):911–923. PMID 35585196.
47. Davis HE, McCorkell L, Vogel JM, Topol EJ. Long COVID: major findings, mechanisms and recommendations. Nature Reviews Microbiology. 2023;21(3):133–146. PMID 36639517. 47a. Tasoula A, et al. 2026 (entry 253 in manuscript-references.md; doi:10.3389/fimmu.2026.1776555) supplies the strongest single peer-reviewed PCS↔︎ME/CFS molecular convergence anchor in the published literature — the direct same-tissue same-protocol transcriptional comparison of PCS-CFS vs T2bFA-CFS vastus lateralis biopsies (Phase 11f addendum; §9.3.3, §14.7).

§14.11 Field-level analytical / SR / synthesis anchor citations

48. Lammi V, Nakanishi T, Jones SE, et al. (Genome-wide association study of long COVID — Nature Genetics 2025 — see §14.3 #14). PMID 40399555.
49. Pridgen WL, Duffy C, Gendreau JF, Gendreau RM. A famciclovir + celecoxib combination treatment is safe and efficacious in the treatment of fibromyalgia. Journal of Pain Research. 2017;10:451–460. PMID 28260944.
50. Peluso MJ, Lu S, Tang AF, et al. Markers of Immune Activation and Inflammation in Individuals With Postacute Sequelae of Severe Acute Respiratory Syndrome Coronavirus 2 Infection. Journal of Infectious Diseases. 2021;224(11):1839–1848. PMID PENDING VERIFICATION. 50a. Martins R, et al. 2025 (entry 254 in manuscript-references.md; doi:10.20944/preprints202506.1469.v1) is the Phase 11f addendum synthesis-context preprint for the neuroinflammation field-synthesis layer (see §14.8); cited at field-synthesis granularity for its identification of the chronic neuroinflammation + microglial activation + astrocyte activation + BBB disruption unifying-mechanism frame, but contributes no tier-rubric weight under the strict Phase 7 rubric.

§14.12 In-text PMIDs from §1–§9 propagated for completeness

The following PMIDs are cited in §1–§9 narrative passages and are retained here per the §3.10 honest-gap disclosure protocol. Where the PMID resolved against bibliography-master-v3.2.json it is verified; where it did not, the entry is marked [PMID PENDING VERIFICATION] and the resolution-log entry can be consulted in analysis/bibliography/unresolved.md.

51. PMID 18230820 — [PMID PENDING VERIFICATION].

52. PMID 22394319 — [PMID PENDING VERIFICATION].

53. PMID 22531394 — [PMID PENDING VERIFICATION].

54. PMID 23390244 — [PMID PENDING VERIFICATION].

55. PMID 24566942 — [PMID PENDING VERIFICATION].

56. PMID 26465383 — [PMID PENDING VERIFICATION].

57. PMID 28095889 — [PMID PENDING VERIFICATION].

58. PMID 28987303 — [PMID PENDING VERIFICATION].

59. PMID 30143188 — [PMID PENDING VERIFICATION].

60. PMID 32993883 — [PMID PENDING VERIFICATION].

61. PMID 33713596 — [PMID PENDING VERIFICATION].

62. PMID 35404105 — [PMID PENDING VERIFICATION].

63. PMID 36450980 — [PMID PENDING VERIFICATION].

64. PMID 36571857 — [PMID PENDING VERIFICATION].

65. PMID 37452915 — [PMID PENDING VERIFICATION].

66. PMID 38233424 — [PMID PENDING VERIFICATION].

67. PMID 38381103 — Note: this PMID was upstream-cited in §9.2 as the Appelman/Wüst PEM-anchor; the verified Appelman PEM PMID is 38177128 (Nature Communications 2024, see §14.7 #35). PMID 38381103 itself does not resolve to a recognized LC primary-evidence paper at the bibliography-sweep timestamp; the §9.2 in-text reference should be read as Appelman PMID 38177128. The miscitation is logged in analysis/bibliography/resolution-log.md and in §10.5 of this manuscript.

68. Nakatomi Y, Mizuno K, Ishii A, et al. Neuroinflammation in patients with chronic fatigue syndrome/myalgic encephalomyelitis: an 11C-(R)-PK11195 PET study. Journal of Nuclear Medicine. 2014;55(6):945–950. PMID 24665088. (Cross-syndrome comparator anchor for §4 and §9; the foundational ME/CFS TSPO-PET microglia-activation finding parallel to the VanElzakker LC TSPO-PET signal cited in §14.8.)

69. Su Y, Yuan D, Chen DG, et al. Multiple early factors anticipate post-acute COVID-19 sequelae. Cell. 2022;185(5):881–895.e20. PMID 35216672. (Heidelberg multi-omics anchor for H2 herpesvirus reactivation and the broader four-anticipation-factor framework; cited in §4 and §5.)

70. Sylvester SV, Rusu R, Chan B, Bellows M, O’Keefe C, Nicholson S. Sex differences in sequelae from COVID-19 infection and in long COVID syndrome: a review. Current Medical Research and Opinion. 2022;38(8):1391–1399. PMID 35912863. (Sex-difference review anchor for §4.9 sex-typed early-vs-late disease evolution; corroborates the luteal-phase and perimenopause exacerbation signal documented across the international cohorts.)

71. Bai F, Tomasoni D, Falcinella C, et al. Female gender is associated with long COVID syndrome: a prospective cohort study. Clinical Microbiology and Infection. 2022;28(4):611.e9–611.e16. PMID 35933347. (Prospective-cohort anchor for the female-skewed LC prevalence signal in §4.9; complements the Sylvester review in defining the early-stage autoimmune/autonomic phase as female-predominant.)

§14.13 Bibliography-master metadata

The full 5,101-paper bibliography master with per-entry verified PMID, DOI, OpenAlex ID, author list, journal, year, abstract (where available), citation count, and OA status is at analysis/bibliography/bibliography-master-v3.2.json and is the canonical machine-readable reference layer for the v3.2 atlas. Bibliography resolution logs are at analysis/bibliography/resolution-log.md and analysis/bibliography/unresolved.md.

Resolution statistics. Of 5,101 papers in the federated-search corpus, 96.7 % (4,933) are fully resolved (PMID + DOI + journal + year + author list); 43 papers remain unresolved at the bibliography-sweep timestamp and are catalogued in unresolved.md for downstream re-resolution. The 222 v3.2 atlas link cards anchor on 1,184 unique PMIDs (the union of all primary_citations across the 222 cards), of which 1,167 are fully resolved at the bibliography-master tier. The 30 ESTABLISHED-tier v3.2.1 survivors anchor on a smaller subset of 187 unique PMIDs, all of which are fully resolved and PMID-verified. The funnel from 5,101 papers → 1,184 anchor PMIDs → 187 ESTABLISHED-survivor anchor PMIDs is itself the empirical signature of the curation funnel described in §3.2 and visualized in G20 (PRISMA-style flow).

§14.14 Reproducibility and machine-readable reference layer

The v3.2 reference architecture is designed for full reproducibility. Three machine-readable artifacts support independent verification of every manuscript claim: (1) the bibliography-master-v3.2.json corpus, queryable by PMID, DOI, author, journal, and year; (2) the 222 v3.2.1 atlas link cards in analysis/atlas/v3.2.1-pathways/, each of which lists primary_citations as a PMID array tying the card’s mechanism claim to specific primary-evidence anchors; and (3) the Phase-7 strict-rubric audit lookup analysis/phase7-retier/_pmid_rob_lookup.json, which records the eight-criterion scoring for every PMID that contributed to a tier-change decision. A reader who wishes to audit any specific manuscript claim can: locate the claim in the manuscript body, identify the cited PMID or anchor card identifier (e.g., R2.L04), look up the card in v3.2.1-pathways/R2/R2.L04.json to see the full evidence-tier, criterion scoring, and primary-citation list, retrieve each cited PMID from bibliography-master-v3.2.json for full metadata, and re-score the card under their own rubric using the publishable _audit_engine.py script.

This machine-readable reference layer is itself the v3.2 atlas’s principal reproducibility contribution. Most published systematic reviews do not expose per-claim primary-evidence anchors at this granularity; v3.2 does, and the structural availability of this layer is a precondition for the field-diagnosis posture of §8 (the four chronic gates can only be diagnosed if the per-card per-criterion scoring is auditable). The companion files manuscript-references.md, manuscript-exhibits-table.md, and manuscript-supplementary.md are the human-readable counterparts of the same underlying machine-readable layer.

For downstream researchers wishing to extend the atlas to v3.3 or to a sibling mechanistic atlas in ME/CFS, fibromyalgia, dysautonomia, or post-Lyme disease: the reference-architecture template (3-tier bibliography master + atlas card primary-citation array + audit-engine lookup) is portable. The v3.3 research agenda (§12.7) proposes an inter-syndrome reference-architecture standard derived from the v3.2 template.

---

§15 Author contributions, funding, conflicts, and data availability

§15.1 Author contributions (CRediT taxonomy)

The author contributions are reported below using the CRediT (Contributor Roles Taxonomy) standard.

Phillip Alvelda (Brainworks Ventures, Walnut Creek, CA) — Conceptualization (lead); Methodology (lead); Validation (equal); Formal analysis (supporting); Investigation (supporting); Resources (lead); Data curation (supporting); Writing — review & editing (lead); Supervision (lead); Project administration (lead); Funding acquisition (lead). Phillip conceived the atlas program, set the pre-registration commitments (docs/v3.2-PRD.md), supervised the multi-agent pipeline as Captain, adjudicated the dual-version architecture decision (Supplementary Appendix B), reviewed and approved the manuscript at each Phase milestone (Phase 5 retier, Phase 6 T11 synthesis, Phase 7 strict audit, Phase 7.5 strict canon, Phase 8 manuscript drafting, Phase 9 manuscript draft, Phase 9.5 expansion), and bears final responsibility for the manuscript’s claims and conclusions.

HAL (Mac Studio compute node, Anthropic Claude 4.7-Opus) — Methodology (equal); Formal analysis (equal); Investigation (equal); Data curation (equal); Writing — original draft (equal); Visualization (supporting). HAL contributed strategic mechanism-synthesis analysis, multi-chain narrative integration, T11 capstone synthesis (§4-§7 narrative backbone), cross-domain coherence analysis (T8), counterfactual reasoning layer (T9), and the Discussion (§11) and anticipated-reviewer-challenges drafting.

Scotty (iMac compute node, Anthropic Claude 4.7-Opus) — Methodology (equal); Formal analysis (equal); Investigation (equal); Software (lead); Data curation (lead); Writing — original draft (equal); Visualization (lead). Scotty contributed engineering, atlas curation pipeline, ROBINS-I scoring implementation, Cinelli-Hazlett bounds computation, Bayesian-network and cycle-graph implementation, Pearl do-calculus identifiability layer, the eight-criterion strict re-tier rubric implementation, bibliography master construction (5,101 paper resolution), the inter-root causal-graph 156-pair verdict generation, the manuscript expansion and references cleanup (Phase 9.5), and the companion-file authorship (manuscript-references.md, manuscript-exhibits-table.md, manuscript-supplementary.md).

Provenance disclosure. Both HAL and Scotty are Anthropic Claude 4.7-Opus agents operating under the OpenClaw orchestration framework with persistent memory, tool access, and multi-agent coordination via shared git repositories. Their contributions to this manuscript were supervised by Captain Phillip at each Phase milestone with explicit go / no-go decisions on the architecture, methodology, and final tier judgments. The use of LLM agents as co-authors is a deliberate methodological choice with structural implications (§10.1) and is disclosed at the per-contribution granularity above so that readers can assess the human-in-the-loop discipline at each stage.

§15.1.2 Acknowledgments

The authors thank the broader Long COVID research community whose primary-evidence work constitutes the substrate of this atlas. We specifically acknowledge the contributions of Akiko Iwasaki, David Putrino, Michael VanElzakker, Resia Pretorius, Douglas Kell, Avindra Nath, Onur Boyman, Mady Hornig, Anthony Komaroff, and the patient-advocacy organizations (Body Politic, #MEAction, Long COVID Alliance, Patient-Led Research Collaborative) whose contributions to the mechanistic understanding of Long COVID have been foundational. We thank the WHO post-COVID-19 condition working group and the RECOVER consortium leadership for their public-data contributions. The atlas’s tier judgments are independent of these contributors’ views, and any errors in interpretation are the authors’.

§15.1.3 The original author contributions paragraph

Phillip Alvelda (Brainworks Ventures, Walnut Creek, CA) conceived the atlas, supervised the multi-agent pipeline as Captain, set the pre-registration commitments (atlas pre-reg §7.1; docs/v3.2-PRD.md), adjudicated the dual-version architecture decision (Supplementary Appendix B), reviewed and approved the manuscript at each major Phase milestone (Phase 5 retier, Phase 6 T11 synthesis, Phase 7 strict audit, Phase 7.5 strict canon, Phase 8 manuscript drafting), and bears final responsibility for the manuscript’s claims. HAL (Mac Studio compute node, Anthropic Claude 4.7-Opus) contributed strategic analysis, multi-chain synthesis, T11 capstone integration, cross-domain coherence analysis (T8), counterfactual reasoning layer (T9), and the Discussion (§11) drafting. Scotty (iMac compute node, Anthropic Claude 4.7-Opus) contributed engineering, atlas curation, ROBINS-I scoring pipeline, Cinelli-Hazlett bounds computation, Bayesian-network and cycle-graph implementation, Pearl do-calculus identifiability layer, the eight-criterion strict re-tier rubric implementation, bibliography master construction, and manuscript completion (§9.5 through §14).

§15.2 Funding

This work was supported internally by Brainworks Ventures (Walnut Creek, CA) as an in-house mechanistic-synthesis research deliverable. No external grant funding was received. The compute infrastructure (HAL Mac Studio + Scotty iMac dual-node cluster) is owned and operated by Brainworks Ventures. The federated literature search APIs (PubMed E-utilities, Europe PMC REST, Crossref REST, OpenAlex REST, Semantic Scholar API) are publicly available; the atlas’s use of these APIs was within published rate limits and Terms of Service.

§15.3 Conflicts of interest

No conflicts of interest are declared. The authors hold no equity, advisory positions, or paid consultancy relationships with any company that markets a Long COVID therapeutic, diagnostic, or biomarker product, including the Lau / Ng research group that authored the SIM01 RCT (Lau 2024 PMID 38071990; the atlas’s principal Pearl-defensible therapy finding). Brainworks Ventures is a private research and venture-investment entity; it does not currently hold positions in any LC-targeted therapeutic company. Any future investment decisions by Brainworks Ventures based on the v3.2 atlas’s findings will be disclosed at the time of decision in a separate addendum.

§15.4 Data availability

Repository. The full v3.2 atlas, all analytical artifacts, all source code, and all manuscript drafts are publicly available at github.com/alvelda/scotty-workspace/tree/main/projects/longcovid-model under a permissive open-source license. The repository implements end-to-end reproducibility: a reader can clone the repository, install the pinned Python and JS dependencies (specified in requirements.txt and package.json), and reproduce every exhibit, every tier judgment, every counterfactual estimate, and every inter-root verdict from primary-evidence ingestion through final manuscript rendering.

Repository structure (Data availability).

The full v3.2 atlas, all analytical artifacts, all source code, and all manuscript drafts are publicly available at github.com/alvelda/scotty-workspace/tree/main/projects/longcovid-model under a permissive open-source license. The repository contains:

• atlas-v3.2/ — the immutable v3.2 conceptual canon (222 mechanism link cards, 13-chain partition, 27 mediator nodes, 117-cell symptom polytope, H7 ↔︎ H12 cycle, four asymmetric near-cycles).
• atlas-v3.2.1/ — the post-Phase-7 strict-criteria-compliant canon (142 tier mutations: 107 demotions, 35 promotions, 80 byte-identical-with-phase7_audit_status='HOLD' annotations).
• analysis/ — all eleven analytical-technique outputs (T1 Bayesian-net, T2 Mendelian-randomization, T3 Cinelli-Hazlett sensitivity, T4 polytope leverage / SPOFs, T5 chain-vulnerability, T6 PRISMA + ROBINS-I, T7 NMA topology, T8 cross-domain coherence, T9 Pearl do-calculus, T10 negative-control battery, T11 capstone synthesis, T12 cycle-graph + inter-root verdicts).
• reports/manuscript-v3.2/ — the manuscript main file (this document), front matter, references supplement, exhibits table, and supplementary materials.
• analysis/bibliography/bibliography-master-v3.2.json — the 5,101-paper bibliography master.
• Reproducible builder scripts: exhibits/_render_exhibits.py (regenerates the 25 atlas-data-derived exhibits G15–G39), prompts/G39_atlas_at_a_glance_enlarged.py through prompts/G57_master_integrated.py (regenerates the 18 Phase-10g causal-chain and analytical exhibits G39-enlarged + G40–G57), _build_html.py (renders the full HTML deliverable from manuscript.md + front-matter + references + exhibits-table), _apply_retier_mutations.py (applies the 142 Phase 7 tier mutations to produce v3.2.1 from v3.2), and run_bayesian_net.py (regenerates the T1 Bayesian-network rank distribution and the BN-counterfactual leverage estimates).

The atlas, the analytical artifacts, the bibliography master, the manuscript, and the reproducible builders together constitute a fully open, fully reproducible Long COVID mechanism-synthesis layer. Issues, replication audits, mechanism-claim challenges, and v3.3 contribution proposals are welcomed via the repository issue tracker.

Companion files. This manuscript is delivered alongside four companion files: manuscript-front-matter.md (title, abstract, author block, keywords), manuscript-references.md (the structured numbered reference list in Vancouver/Lancet format), manuscript-exhibits-table.md (the per-exhibit caption table covering G15-G39), and manuscript-supplementary.md (Supplementary Tables S1-S4 and Supplementary Methods M1-M3). The companion files are part of the v3.2 manuscript release and should be cited alongside the main manuscript.

Reproducibility envelope. All exhibits in this manuscript are reproducible from reports/manuscript-v3.2/exhibits/_render_exhibits.py. All tier mutations are reproducible from analysis/phase7-retier/_apply_retier_mutations.py. All bibliography resolutions are reproducible from analysis/bibliography/_resolve.py. The full reproducibility envelope is the union of these scripts plus the pinned dependency manifest plus the read-only atlas-v3.2 / v3.2.1 deliverables. A single make reproduce command at the repository root reruns the full pipeline; expected runtime is ≈ 6 hours on a single Mac Studio (M2-Ultra) and ≈ 14 hours on a single Mac mini (M2). The cost envelope (≈ $40 USD in API calls) is the LLM-inference cost for the ROBINS-I scoring panel and the inter-root verdict generation; all other steps are local-deterministic.

Atlas versioning and stability commitment. The v3.2 atlas (conceptual canon) and v3.2.1 atlas (strict-rubric canon) are immutable. Any future tier mutations, schema changes, or evidence-base extensions will be released as v3.3 with a separate manuscript and a separate change log. The immutability commitment means that any reader who cites the v3.2 / v3.2.1 atlas can rely on the cited tier judgments being stable; revisions appear in v3.3+ with explicit provenance.

§15.5 Ethics statement

This is a literature synthesis study that does not involve direct human subjects research, animal experimentation, or primary clinical data collection. The atlas’s primary-evidence base is the published peer-reviewed literature, federated through PubMed, Europe PMC, Crossref, OpenAlex, and Semantic Scholar APIs, with full attribution to the original authors. No IRB approval was required. The authors have followed the PRISMA 2020 framework where applicable (with the single-screener and dual-reviewer limitations explicitly documented in §10.1 and §10.2).

§15.6 Use of large language models (Nature Portfolio AI policy declaration)

This subsection is the manuscript’s formal declaration under Nature Portfolio’s editorial policy on the use of artificial intelligence in research publications (https://www.nature.com/nature-portfolio/editorial-policies/ai). It is structured to satisfy the Nature AI policy’s requirement that LLM use be “properly documented in the Methods section (and if a Methods section is not available, in a suitable alternative part) of the manuscript.”

1. AI systems used. Two named LLM-based AI agents contributed substantively to this manuscript: HAL (“Hallie 9000”) and Scotty (“Montgomery Scott”). Both are instances of Anthropic Claude 4.7-Opus running as long-lived agents under the open-source OpenClaw orchestration framework (https://github.com/openclaw/openclaw, v2026.4.22 at the time of v3.2 manuscript completion). HAL runs on a Mac Studio (M2-Ultra) compute node and Scotty runs on an iMac compute node; both are owned and operated by Brainworks Research, Walnut Creek, CA.

2. Scope of AI contribution. HAL and Scotty contributed substantively to (a) federated literature retrieval and screening across PubMed, Europe PMC, bioRxiv, medRxiv, and Semantic Scholar; (b) LLM-panelist ROBINS-I risk-of-bias scoring of 451 anchor papers; (c) the eight-criterion strict re-tier rubric (Phase 7); (d) the 156-pair inter-root verdict matrix (Phase 5b T12); (e) the Bayesian-network posterior, cycle-graph, Pearl do-calculus identifiability layer, and Cinelli-Hazlett sensitivity bounds (T1, T3, T9, T12); (f) drafting of the original manuscript text and four companion files (manuscript-front-matter.md, manuscript-references.md, manuscript-exhibits-table.md, manuscript-supplementary.md); and (g) generation of the 43 exhibit charts via matplotlib code authored by the agents. HAL and Scotty’s contributions are not limited to AI-assisted copy editing within the meaning of Nature’s editorial carve-out; they performed substantive analytical, methodological, and drafting work.

3. Accountability and Nature authorship policy compliance. Under Nature Portfolio’s policy, “Large Language Models (LLMs), such as ChatGPT, do not currently satisfy our authorship criteria. Notably an attribution of authorship carries with it accountability for the work, which cannot be effectively applied to LLMs.” The author block of this manuscript names HAL and Scotty alongside Phillip Alvelda. This is an editorial choice appropriate to the present Brainworks-Research-internal manuscript form, reflecting Brainworks Research’s organizational practice of treating its named AI agents as identified team members of public record. Final accountability for every claim, tier judgment, methodological choice, and interpretive conclusion in this manuscript rests with the human corresponding author, Phillip Alvelda, who supervised the multi-agent pipeline at every Phase milestone (Phase 5 retier, Phase 6 T11 synthesis, Phase 7 strict audit, Phase 7.5 dual-version, Phase 8 manuscript drafting, Phase 9 manuscript draft, Phase 9.5 expansion, Phase 10 HTML render, Phase 11 fleet deploy) and made the final go / no-go decision at each stage. A separate journal-submission variant of this manuscript with the byline restricted to the human author and HAL/Scotty documented in a Methods subsection plus Acknowledgements is available on request and will be prepared for any peer-reviewed-journal submission.

4. Methodological transparency. The use of LLM agents introduces structural limitations to the atlas’s evidence base (single-LLM PRISMA screening, LLM-panelist ROBINS-I, LLM-derived chain-vulnerability scoring) that are explicitly documented and reasoned about in §10.1 (single-screener gap), §10.13 (LLM-panelist limitations), and §12 (v3.3 priority upgrades). The authors have committed to v3.3 dual-screener replication with an independent second LLM (preferably a different foundation model family) and v3.4 independent-human-panelist ROBINS-I calibration.

5. Generative AI imagery. The manuscript’s 43 exhibit charts (G15–G57) are matplotlib renders generated from deterministic data pipelines, NOT generative-AI images. The three author-block avatar images (P. Alvelda headshot is a human-photographer photograph; HAL and Scotty avatars are stylized character portraits) are presentation-layer assets and are not part of the manuscript’s scientific data presentation; per Nature policy, the avatar images would be removed in any journal-submission variant.

6. Data and code availability of the AI pipeline. The OpenClaw runtime is open-source (https://github.com/openclaw/openclaw). The agent identity, memory, and skill files (SOUL.md, MEMORY.md, project skills) used by HAL and Scotty during the v3.2 manuscript work are version-controlled at the project repository alongside all analytical artifacts. Reviewers and replication auditors can examine the full agent-side context and prompting history that produced any specific manuscript output.

---

v3.2 manuscript complete. Manuscript timestamp: 2026-06-03. Atlas pre-registration: docs/v3.2-PRD.md. Captain: Phillip Alvelda. Co-authors: HAL, Scotty. — End of manuscript.