H-1 - SARS-CoV-2 first infected humans via zoonotic spillover at the Huanan market
Link to original
H-5 - SARS-CoV-2 first infected humans through a research-related incident at the WIV
Link to original
H-6 - The first human infection arose by neither route listed here
Catch-all so that “exactly one member is true” holds for this cluster. It is not a bare “other”: it names the concrete escapes the curated literature leaves open — spillover outside the Huanan market (upstream in the same wildlife trade, or geographically distant with later introduction into Wuhan), and a research-related incident at a Wuhan institution other than the WIV. Both are compatible with the whole curated observation set, because every market-side observation (O-1..O-5) is sampled at Huanan and every research-side observation (O-11, O-12) concerns the WIV/EcoHealth DEFUSE consortium specifically; neither dataset can see outside its own frame.
Link to original
Why this carving
The three surviving hypotheses do not form three competing answers to one question, so the carving decision was not free.
H-1 and H-5 are genuinely exclusive and both answer the same dimension — where and how the index infection occurred. Their step-3 question fields agree, and here the labels match the statements.
H-3 (natural evolution/recombination in the wild Rhinolophus sarbecovirus reservoir) is not exclusive with either of them. A virus can evolve naturally in bats and spill over at the Huanan market; it can equally evolve naturally in bats, be collected into a laboratory, and escape. H-3 is therefore compatible with every member of this cluster and cannot sit in it as a competing answer — putting it in would break exclusivity and make the cluster stop being a probability space.
The tempting second cluster was rejected as evidence double-counting. The obvious alternative is a second cluster on the evolutionary origin dimension: {H-3 “natural evolution”, + a constructed complement “the genome was deliberately manipulated in a laboratory”, + a residual}. That complement is, in this evidence base, H-5’s own content restated: H-5 explicitly reads “potentially involving an engineered furin cleavage site of the kind proposed in DEFUSE”, and the three observations that would discriminate the constructed cluster — O-10 (no furin cleavage site anywhere in the BANAL clade), O-11 (DEFUSE proposed inserting exactly such sites at the WIV), O-12 (DARPA rejected DEFUSE) — are the same three that discriminate H-5 from H-1 in this cluster, pushing in the same direction in both. Two clusters that are alternative slicings of one sub-question double-count that evidence at step 8, and rule 3 states plainly that no depends_on note repairs it. So the second cluster was not built.
H-3 was therefore dropped as a straggler that clusters nowhere — not because it is off-topic (it is not), but because it is a background claim rather than a competing answer: it is entailed by H-1, consistent with H-5, and its distinguishing content (natural versus engineered genome) already lives inside H-5’s statement, where O-10/O-11/O-12 will score it exactly once. The genomic observations O-6 to O-9 are not orphaned by this: a naturally circulating, human-ACE2-binding sarbecovirus reservoir bears directly on H-1 versus H-5 within this cluster, since it speaks to whether a natural progenitor was available without laboratory involvement. See agent-notes/structure.md and the problem log for the pipeline tension this exposed (the ≥2-members rule leaves no non-distorting home for a true-but-non-competing hypothesis).
Grain. The two named members sit at comparable specificity — each names a place plus a mechanism (Huanan market via live wildlife; WIV via a research-related incident) — so neither is an over-narrow member competing against a broad one. H-6 carries everything outside both frames, which matters here because the curated evidence is doubly frame-bound: the market observations sample only Huanan, and the research observations concern only the WIV/EcoHealth DEFUSE consortium.
Not exhaustive by construction, hence the residual: the two named members are the two positions this literature argues, not a logical partition of the space of first infections.
Prior
# HC-1 prior — route of the first human SARS-CoV-2 infection. Members in HC-1.hypotheses order:
# [H-1 Huanan market spillover, H-5 WIV research-related incident, H-6 residual].
#
# EVIDENCE SELECTION (rule 5): NOTHING is marked used_for_prior. All 9 inbound edges sit in one of
# 4 correlation groups (CG-1..CG-4) and marking is all-or-none per group, so there is no way to pull
# a base-rate fact in without also removing a case-specific discriminating one. The candidate was
# CG-2: O-4 (47,381 live animals of 38 species at Wuhan markets, incl. susceptible raccoon dogs) is a
# genuine precondition/base-rate fact, but its group-mate O-5 (no bats or pangolins on sale) is
# explicitly discriminating — the edge note itself says it "mildly favours H-5/H-6". Marking CG-2
# would delete that discrimination from step 8 to buy a base rate the prior can carry as a reasoned
# factor instead. Same verdict for CG-1, CG-3, CG-4, which are all case-specific. So: pure outside-view
# prior, built from emergence-event base rates only, and step 8 prices all 9 edges.
# `no_observation_arguments.py --cluster HC-1` returns none (A-5 is skipped by the helper), so no
# argument input reaches this block (rule 6 is vacuous here).
# ---------------------------------------------------------------------------------------------
# Anchor. Take "the virus reached its first human host by natural zoonotic transmission somewhere in
# the wildlife/livestock chain" as the unit of relative weight. Reference class: novel human-pathogenic
# virus emergence events of the last century (SARS-1, MERS, HIV-1/2, Ebola, Nipah, Hendra, influenza
# reassortants), which are overwhelmingly natural.
w_natural_route_total = 1.0 # unit anchor; scale is free (rule 3)
# Odds that an emergence event is research-related rather than natural, before any Wuhan-specific fact.
# Reference class: documented laboratory-origin human outbreaks that spread in the community — the 1977
# H1N1 re-emergence, Birmingham smallpox 1978, the 2004 Beijing SARS-1 escapes (4 separate escape events
# after the natural outbreak ended). For this value to be right, roughly 1 in ~30 novel-agent community
# outbreaks would have to trace to a research incident. Lab-acquired *infections* are far commoner than
# this; the rarity is in seeding onward community spread.
r_research_vs_natural_base = 0.035
# Visible adjustment for the setting, not for any case-specific observation: the outbreak city hosts the
# world's largest bat-sarbecovirus collection and sampling programme. Under a research-incident account
# the coincidence of place is expected; under a natural account it is a coincidence. For this factor to
# be right, the geographic coincidence would have to be worth ~one bit of evidence — a coarse
# order-of-magnitude judgement, and the single most arguable number in this block.
uplift_coincident_lab_city = 5.0
w_research_route_total = r_research_vs_natural_base * uplift_coincident_lab_city * w_natural_route_total
# ---------------------------------------------------------------------------------------------
# H-1 — zoonotic spillover at the Huanan market via live wildlife sold there.
# What would have to be true: a susceptible live animal carrying an ancestral sarbecovirus was physically
# present at Huanan and infected a human *there*, i.e. the market is the site of the index infection, not
# merely an amplifier. Basis: the natural-emergence anchor above, times the share of natural spillovers
# whose first human case occurs at the retail node of the chain rather than upstream (farm, trapper,
# trader, transport) — SARS-1's own index infections were largely at market/restaurant/handler nodes,
# which is why this share is sizeable rather than small. What it does NOT cover: any of the market's own
# evidence (case geolocation, environmental swab clustering) — those are step 8's. Confidence: moderate;
# this is the member whose weight will move most once CG-1/CG-2 are priced.
p_index_case_at_retail_market_given_natural = 0.40 # retail node is the highest-human-contact link, but one of ~4-5 plausible links
w_h1 = w_natural_route_total * p_index_case_at_retail_market_given_natural
# ---------------------------------------------------------------------------------------------
# H-5 — a research-related incident at the WIV, possibly involving an engineered furin cleavage site.
# What would have to be true: a WIV worker was infected during collection, culture, passage or
# engineering work and carried the virus into the city. Basis: the lab-outbreak reference class above,
# uplifted for the coincidence of city, then narrowed to the WIV specifically among Wuhan institutions
# that handle bat coronaviruses (WIV, Wuhan CDC, university and hospital labs). What it does NOT cover:
# DEFUSE (O-11/O-12) or the furin-site genomics (O-10) — deliberately left unpriced here so step 8 prices
# them once. Confidence: this is the least confident of the three; both factors are order-of-magnitude
# judgements and a reader who rejects uplift_coincident_lab_city should re-run with it set to 1.
p_wiv_given_research_incident = 0.70 # WIV ran by far the largest sarbecovirus programme in the city; other Wuhan institutions handled such material at much lower volume
w_h5 = w_research_route_total * p_wiv_given_research_incident
# ---------------------------------------------------------------------------------------------
# H-6 — residual: neither listed route. Concretely this looks like (a) a natural spillover to a farmer,
# trapper, trader or transporter upstream of Huanan, or at one of Wuhan's other markets, with the virus
# then carried into the Huanan crowd and amplified there; (b) a spillover hundreds of km away with silent
# transmission into Wuhan before detection; or (c) a research-related incident at a non-WIV Wuhan
# institution. Argued in its own right (rule 4), not as a leftover. Basis: the same emergence reference
# class — in most historically resolved zoonoses (SARS-1, MERS, HIV, Ebola) the index infection turned out
# NOT to be at the venue where the outbreak was first noticed, and the venue of first detection is a
# strongly biased estimator of the venue of first infection. What it does not cover: it inherits the
# uncertainty of the two named members' own decompositions. Confidence: high that this deserves a large
# weight; the curated evidence base is doubly frame-bound (only Huanan sampled, only WIV/DEFUSE examined),
# so an unmodelled route is exactly the kind of answer this graph could not see.
w_h6_natural_offmarket_leg = 0.50 # relative to the unit natural anchor: upstream-chain or distant-spillover-with-silent-introduction routes, collectively about as likely as the market-index route
w_h6_research_nonwiv_leg = 0.06 # a research incident at a Wuhan institution other than the WIV; small because the WIV dominates the relevant material volume
w_h6 = w_h6_natural_offmarket_leg + w_h6_research_nonwiv_leg
prior("HC-1", [w_h1, w_h5, w_h6])