Did SARS-CoV-2 first infect humans through natural zoonotic spillover (e.g. via the wildlife trade / Huanan Seafood Market) or through a research-related incident (a lab leak)?
1. The answer
Natural zoonotic spillover on this evidence base — but this analysis does not settle it, and settles the Huanan-market version least of all. The KB’s one cluster, HC-1 - Route of the first human SARS-CoV-2 infection, returns (baseline run run.py <analysis-dir>, 4 evidence blocks, byte-identical on re-run) a posterior of H-1 - SARS-CoV-2 first infected humans via zoonotic spillover at the Huanan market 0.4952 · H-5 - SARS-CoV-2 first infected humans through a research-related incident at the WIV 0.0809 · H-6 - The first human infection arose by neither route listed here 0.4238, from a prior of [0.3695, 0.1132, 0.5173]. H-6 is not a third answer but a bag holding both an off-market natural route and a non-WIV research route, weighted 0.50 against 0.06 in the prior — read through that weighting the model leans natural. But the 42.4% residual is, per Analysis of HC-1 - Route of the first human SARS-CoV-2 infection, structural, not arithmetic: every market-side observation was sampled at Huanan and every research-side one concerns WIV/DEFUSE, so no likelihood can discriminate against either residual leg — a fact about the observation set, not the world.
Two reasons the number is weaker than it looks. First, H-5’s 8.1% is not a finding: it is the step-7 prior surviving weak evidence, and that prior rests on one Fermi factor, uplift_coincident_lab_city = 5.0, with no reference class behind it. Run --set HC-1:uplift_coincident_lab_city=1.0 takes H-5 to 0.0179; runs --set CG-4:lik_defuse_H5=3.0 → 0.1725 and --set CG-1:lik_swabs_H5=1.0 → 0.1469 roughly double it. A quantity moving tenfold under defensible re-settings of one unsourced multiplier is not a measurement. Second, the N=5 cut removed an entire evidence class: case-geolocation and epidemiology are absent from the model altogether — S-13 - Pekar, Worobey, Wertheim et al. 2022 Science - lineage A and B imply at least two independent zoonotic spillovers sat at the pool’s top usefulness (4.5) and was cut, as was S-3 - Crits-Christoph, Débarre, Worobey et al. 2024 Cell - Susceptible wildlife DNA co-located with SARS-CoV-2 in Huanan market samples, leaving D-2 - Huanan market China CDC environmental and animal swab raw metagenomic dataset priced through one reading only.
Outside the number. The model sees no route outside Huanan or the WIV; it does not separate unmodified-collection escape from engineered escape; its prior took zero evidence and zero argument input; and the counterargument nobody curated had an incentive to make — ascertainment bias toward market-linked cases, S-5 - Bloom 2021 MBE - Recovered deleted early-Wuhan sequences suggest ascertainment bias toward market-linked viruses — was cut.
The bet test. I would bet natural-side against research-related at roughly 3:1 to 5:1 — the report’s own judgment, not a model output — less confident than the model’s members imply, because the arm the odds run against is unevidenced here rather than argued down. On the narrower claim that the index infection happened at Huanan rather than upstream, I would bet at no odds: that is the distinction this sampling frame cannot make.
Entry points: Analysis of HC-1 - Route of the first human SARS-CoV-2 infection · HC-1 - Route of the first human SARS-CoV-2 infection · H-6 - The first human infection arose by neither route listed here · CG-1 - HC-1 joint over O-1+O-2 · CG-4 - HC-1 joint over O-11+O-12 · O-1 - 74 of 923 Huanan environmental swabs SARS-CoV-2-positive, concentrated in the live-wildlife section · D-2 - Huanan market China CDC environmental and animal swab raw metagenomic dataset · O-3 - Huanan market closed and disinfected from 1 Jan 2020 with animal sampling only from 18 Jan
2. What the analysis found
The knowledge base has exactly one cluster. HC-1 - Route of the first human SARS-CoV-2 infection is the main question, restated over three exclusive members. The weighing step that normally carries a report of this kind — combining sub-answers in different currencies — therefore does not exist here, and I will not manufacture a cross-cluster argument to fill the space. All of step 10’s judgment goes into one place instead: deciding how much of HC-1’s posterior to believe, and how to map three members onto a two-way question.
Where it landed and what did the updating: four joint likelihood blocks, no lone edges. CG-1 - HC-1 joint over O-1+O-2 (environmental-swab clustering, lik_swabs_H1 = 1.0 vs lik_swabs_H5 = 0.3) and CG-4 - HC-1 joint over O-11+O-12 (DEFUSE, lik_defuse_H5 = 1.0 vs lik_defuse_H1 = 0.40) pull in opposite directions; CG-1 wins, H-5 falls 0.1132 → 0.0809 and H-1 rises 0.3695 → 0.4952. CG-3 - HC-1 joint over O-8+O-9+O-10 nets a pro-H-1 signal (a wild sarbecovirus already binds human ACE2) against a pro-H-5 one (O-10 - None of the three BANAL viruses carries a furin cleavage site) into a single number, burying the latter.
The mapping judgment. H-6 is not “neither answer”; it straddles the question. I treat its 42.4% as mostly natural-side (its prior legs are 0.50 natural-off-market against 0.06 research-non-WIV) but I decline to allocate it numerically, because doing so would import the prior’s leg ratio as if it were a result. That refusal, not an allocation, is what makes part 1’s answer a range rather than a point.
3. What the answer hangs on, and what would change it
Three levers, priced where a variable exists.
uplift_coincident_lab_city = 5.0in HC-1’s prior — the largest single lever, and unsourced. Run--set HC-1:uplift_coincident_lab_city=1.0→ [0.5263, 0.0179, 0.4558]: H-5 collapses by a factor of 4.5. Per the review, HC-1’s prior received zero evidence input and zero argument input (a logged pipeline bug plus all-or-none group marking meant A-5 - DEFUSE documents intent and capability to engineer exactly the furin site SARS-CoV-2 has but absent from relatives, raising lab-origin probability reached nothing), so H-5’s posterior is largely this ungrounded prior surviving.- The CG-3 net.
--set CG-3:lik_banal_H5=1.3→ [0.4713, 0.1247, 0.404]: exposing O-10’s isolable pro-H-5 signal is worth ~50% on H-5. Its true value depends on how densely the sarbecovirus reservoir has been sampled — a quantity no curated source estimates. - The CG-1 anchor.
--set CG-1:lik_swabs_H5=1.0→ [0.4564, 0.1469, 0.3966]. Note the swab clustering rests on O-1 - 74 of 923 Huanan environmental swabs SARS-CoV-2-positive, concentrated in the live-wildlife section and its companion null O-2 - No Huanan animal-origin sample (457 samples, 18 species) tested SARS-CoV-2-positive, both from the single China-CDC-led reading of D-2.
Unpriceable but larger than any of these: H-5 silently bundles unmodified-escape and engineered-escape, which fit O-11/O-12 very differently. H-3 - SARS-CoV-2 originated by natural evolution in a bat sarbecovirus reservoir was dropped at step 4, so that distinction exists nowhere in the graph and has no variable to move.
Missing information, re-ordered by effect on the answer (labels carried from the review): (i) any observation locating early cases independently of Huanan sampling — the S-1/S-7/S-13 class — exists, unread; this is the only item that could compress H-6, and therefore the only one that could turn the range in part 1 into a number. (ii) D-2’s raw reads read against the S-3 reanalysis — exists, unread. (iii) Whether H-5’s two branches are separable on available evidence — unclear. (iv) Raw line lists for the 174 earliest cases — exists, inaccessible. (v) WIV’s pre-2019 collection inventory and the databases taken offline in 2019 — exists, inaccessible. (vi) Contemporaneous wildlife sampling upstream of Huanan — does not exist.
4. What this does not cover
Scope and framing. The question was taken as: by what route did the first human infection occur — not where the outbreak was first amplified or detected, a distinction the curated evidence is systematically bad at making. Curation (agent-notes/curation.md) kept 5 of 17 sources at a trust baseline lowered from 0.8 to 0.5, and the cut removed an evidence class rather than trimming within classes: epidemiology and case geolocation are wholly absent, as is the ascertainment-bias counterargument. The search itself (agent-notes/orientation/) was split into an epi-field and a genomic-institutional slice with explicit cross-slice exclusions, and treated the debate-hub material as discovery only.
Never modelled. The natural-versus-engineered genome dimension (H-3 dropped, H-4 dropped); non-WIV Wuhan institutions except as an undifferentiated prior leg; motivatedness adjustment — S-2 is China-CDC-led and S-15 was surfaced by DRASTIC, sitting at trust 0.72 and 0.80 with no correction for the direction of their incentives. Orphans: the review judges the orphaning of O-3 - Huanan market closed and disinfected from 1 Jan 2020 with animal sampling only from 18 Jan wrong — it discriminates (it is why O-2’s animal-negative result is weak) but only through A-1 - Post-disinfection late animal sampling means the animal-negative result cannot exclude infected animals adjusting a likelihood inside a comment, a path the graph cannot represent. O-6, O-7 and the unassessed orphan A-3 are correctly out. No cluster carries an unresolved depends_on (there is only one cluster).
What the debate performed rather than settled. Both sides in this corpus argue from the frame their own sampling created; neither curated side produced the observation that would break the other’s frame.
External consensus (external, labelled; entered into no number here). No external_consensus field exists anywhere in this KB — a gap this report notes rather than fills. From my own background knowledge: the published virology mainstream (Worobey/Pekar/Andersen line) holds market zoonosis to be strongly favoured, often stated at odds of tens-to-one or more; US agency assessments split, with FBI (moderate confidence) and DOE (low confidence) leaning research-related and others leaning zoonotic; the Rootclaim/Miller debate judges both ruled for zoonosis at long odds. My bottom line is on the same side but far less confident than the virology mainstream and less confident than the debate judges. What this analysis contributed beyond the consensus is not a rival number — it is the demonstration that a 5-source base which omits the epidemiological evidence class cannot produce a confident verdict on either side, and that the lab-leak arm’s small posterior here is a Fermi prior rather than an evidential defeat. That is a claim about what the field’s confidence is made of, and it is this analysis’s own.