What the analysis says
HC-1 is the analysis’s main question: how did SARS-CoV-2 first reach humans? The space is carved by two nested yes/no questions — was research activity involved in the human introduction, and (if so) had the virus been manipulated in a lab — giving four mutually-exclusive members: [[H-41 - Natural zoonotic spillover with no research involvement]] (a naturally-evolved virus crossing over through ordinary ecological contact, e.g. the wildlife trade), [[H-42 - Research-related leak of an unmodified naturally-evolved virus (zoonotic collection)]] (a wild, unmodified virus that reached humans through the act of collecting/handling it), [[H-43 - Research-related origin of a laboratory-manipulated virus (engineered lab leak)]] (a virus shaped by lab work — gain-of-function, synthesis, chimera-building, or serial passage — then leaked), and [[H-44 - The origin is something not listed here]] (a residual for any history the three miss, e.g. a lab-modified virus that then spilled over naturally).
The prior landed at [0.809, 0.120, 0.047, 0.024] — roughly 81% natural, a research-related bloc (H-42 + H-43) near 17%, of which the engineered member held under 5%. That prior is built as a location-free base rate (global odds over emergence mechanisms, where natural zoonosis dominates because SARS-1, MERS and essentially every recent novel-pathogen pandemic was natural) multiplied by named circumstantial factors that have no object-level evidence attached and so can only live in the prior. The dominant circumstantial factor is f_wuhan — that the outbreak began in the city housing the world’s leading bat-coronavirus lab — set to a modest 5.0 against a stated defensible range of roughly 2 to 80.
Step 8’s discriminating evidence then moved the prior to a posterior of [0.776, 0.162, 0.043, 0.019]. The notable feature is which member rose: the whole research bloc gained (from ~17% to ~20%), but almost all of the gain went to H-42, not H-43. Three signals lifted the lab bloc as a whole — the DEFUSE proposal ([[CG-1 - HC-1 joint over O-35+O-36+O-37+O-38]], high trust t=0.90, anchored on H-43 with a means-motive-opportunity reading of a document written in Wuhan proposing exactly the manipulations later debated), the FBI/DOE lab-lean intelligence ([[CG-2 - HC-1 joint over O-47+O-48+O-49+O-52+O-53]], t=0.65, lifting H-42 and H-43 equally at 1.3x), and the nationwide negative host search ([[E-59 - O-27 × HC-1 — nationwide search found no intermediate host]], which weakly favours the lab bloc because a trade-chain zoonosis “should” have left a findable intermediate). Meanwhile the genome and comparative-genetics evidence pushed the engineered member specifically down: [[CG-4 - HC-1 joint over O-58+O-59+O-60+O-61+O-62+O-64+O-65+O-66+O-67+O-74+O-75+O-76+O-77]] and the BANAL closest-relative group [[CG-11 - HC-1 joint over O-68+O-69+O-70+O-71+O-72+O-73]] both price H-43 at ~0.5x the natural bloc, the SARS-CoV-1 live-market precedent [[CG-3 - HC-1 joint over O-54+O-57]] at 0.65x, and the Ebola reservoir-unfound base rate [[E-58 - O-26 × HC-1 — natural outbreak left reservoir unidentified (base rate)]] at 0.8x. Because H-41 and H-42 share an identical natural genome, all of that genome evidence leaves H-42 untouched — so H-42 collected the lab-bloc lift without paying the engineered-genome penalty, which is why it, not H-43, is where the research probability accumulated.
What the model may not capture
The single largest structural doubt is f_wuhan. The block itself flags it as “the single biggest crux of the whole debate,” and the arithmetic bears that out: with f_wuhan set to 1, the remaining circumstantial factors nearly cancel (the cover-up signals A-13/A-14/A-15 against the absence-of-evidence signals A-16/A-28, netting ~0.77), and the research bloc collapses from ~17% to ~4%. So essentially the entire circumstantial lab-tilt of the prior rides on one number that has no evidence-edge, whose defensible range spans a factor of ~40 (2 to 80), and whose true value is precisely the unresolved Worobey/Pekar-vs-Weissman dispute over whether zoonotic-spillover probability tracks raw wildlife-trade volume (tiny in Wuhan) or human population / urbanisation / detectability (large in an 11M hub with a live-mammal market). It is also entangled with HC-3: “why did the outbreak, and the earliest cases, center on Wuhan and its market?” is partly the same question priced there, so there is a live double-counting hazard across clusters. If one thing in this cluster is wrong, it is most likely the magnitude of this factor. Flagged hard, shipped unfixed — the honest fix is a sensitivity band, not a point estimate.
Second, H-41 and H-42 are barely separable. They posit the same natural, unmodified virus and differ only in the route into humans. None of the genome evidence (CG-4, CG-11) or the market/precedent evidence distinguishes them by construction — they are explicitly tied to equal likelihood throughout. The only things that separate them are f_wuhan (attenuated to half-strength on H-42) and DEFUSE-proximity. So the posterior 0.162 on H-42 — now the second-largest member and the bulk of the “lab-leak” probability — rests on genuinely thin discriminators. A reader should treat the H-41/H-42 split as far softer than the two numbers suggest; much of what looks like resolved probability mass is really one weak lever.
Third, a deliberate conservatism worth surfacing: CG-4’s reliability was capped at t=0.25, the minimum trust over its 13 observations, set by the low-trust Bruttel “synthetic fingerprint” source (S-48, trust 0.25) that shares the reference-genome basis with the higher-trust Andersen natural-origin observations (S-46, 0.5). Per the group-trust rule, one low-trust witness in a joint group binds the whole group, so the strong, higher-trust natural-genome signal in CG-4 was discounted to a quarter of its force. The un-capped natural-origin signal is instead carried by CG-11 (BANAL, t=0.80). This is defensible but means the model is leaving genome discrimination power on the table; a cleaner carve would split the high-trust natural reads from the low-trust engineering-signature reads into separate groups (a step-8 re-run, not an edit here).
Fourth, several circumstantial lab factors — WIV biosafety ([[A-13 - Rootclaim- lax WIV biosafety (other coronaviruses apparently escaping) raises the lab-linked hypotheses]]), the database-offline / RaTG13-rename reading (A-14), China’s non-transparency (A-15) — rest on the low-trust say-so of Rootclaim (S-60, trust 0.30) and Stansifer (S-70, 0.57), whose primary sources (State Department cables, database access logs) were not independently curated in this analysis. The blocks already soften each factor well below its source’s raw multiplier, which is right; but motivated-source blind spots cut both ways, and the counterargument that no source in the curated set had an incentive to make — the mundane-explanation case for each “consciousness of guilt” reading — is under-represented.
Is the true answer even on the list? The residual H-44 holds only ~1.9% posterior. The three members cover the mechanism space well, so I think a small residual is broadly right — but note what an unlisted origin would be: a hybrid history the H-42/H-43 line mis-places (serial passage counted as “not manipulation” is the block’s own example), or a lab-modified virus that escaped into animals and later spilled over naturally. That last one is consequential out of proportion to its probability: it would be a research-related origin that every piece of natural-genome and market evidence would fail to flag, because the genome could look natural and the proximate spillover could look zoonotic. A small number there is not a small risk, and the model’s residual value is a guess, not a measurement.
What would help
- A resolved value (or a defensible narrow band) for
f_wuhan/ P(zoonotic outbreak first detected in Wuhan) — the early-case spatial and ascertainment analyses that would pin it exist but are contested and largely priced in HC-3, and the clean cross-cluster reconciliation does not exist (it is the field’s open crux). This would move HC-1 more than anything else. - Any genuine discriminator between H-41 and H-42 (natural-market spillover vs unmodified-collection leak) — does not exist on current evidence; the two are separated only by route priors.
- The primary sources behind the circumstantial lab factors (State Dept WIV-biosafety cables, the database-offline logs, the RaTG13/BtCoV-4991 rename record) — exists, unread (not curated here).
- Independently verifiable raw data from the nationwide ~80,000-animal host survey, with its sampling frame and timing — exists, inaccessible (China-CDC-gatekept; this is why E-59 is docked to t=0.45).
- A re-carve of CG-4 separating the high-trust Andersen natural-genome reads from the low-trust Bruttel engineering-signature reads, so the natural signal is not capped at t=0.25 — the underlying observations exist, unread as separable groups (a step-8 re-run would realise it).
Confusions and contradictions
The clearest genuine conflict is that two evidence edges point in opposite directions on the same question — “was an animal host ever found?” E-58 (Ebola’s reservoir was never identified, so a missing SARS-CoV-2 reservoir is normal for a natural outbreak) favours H-41; E-59 (a nationwide search found nothing, which the lab members predict and a trade zoonosis does not) favours the lab bloc. The analysis does not paper this over — it prices each separately and lets the runner compose them, explicitly avoiding double-counting the “reservoirs are often unfound” base rate. That is methodologically clean, but it means the net direction of the host-search evidence is a small residual of two opposed, individually-weak signals, and a reader should not take either edge as decisive.
A second, internal-to-the-genome tension: CG-4 asks H-43 to fit two half-patterns at once — an “engineering-signature” half (idealised restriction map, silent mutations at restriction sites, the doubled CGG codon) that needs reverse-genetics assembly, and a “natural-relative” half (suboptimal RBD, no published backbone, live-host O-glycans, an out-of-frame insert) that looks passaged or natural. A clean engineered design struggles to be both, which is part of why the engineered member is only moderately (~0.5x), not decisively, disfavoured — the block keeps it a “live, only-moderately-worse fit” because a trace-free serial-passage build could reproduce much of the natural half. This is real irreducible ambiguity in the genome record, not an artifact of the carving.
Finally, the base-rate disagreement is irreducible at source: the three reference classes feeding the spine (Rootclaim historical incidence ~1/130, Weissman mechanistic ~1/149, Stansifer location-stripped ~1/30) disagree by up to ~4x on the location-free lab:zoonosis odds, and the block deliberately mixes rather than picks. The mixture is the honest move, but it means the prior’s lab-bloc magnitude is a smear over classes that genuinely do not converge, and no evidence in this cluster resolves which class is right.