debate-bayesian-record

Orientation — Debate & Bayesian-analysis record (slice G)

Slice: the primary record of the reasoning itself — the Rootclaim vs. Peter Miller $100,000 judged debate (Nov 2023), the judges’ decisions, and the wider set of independent Bayesian point-estimates the case’s framing summarizes as “six independent analyses … span[ning] 23 orders of magnitude.” 11 source nodes minted, all treated as primary records of their own argument (per this slice’s exception to the normal “no reviews as nodes” rule). No empirical primaries were minted here — only debate/analysis artifacts, per the brief.

Topics

The two debaters’ own primary cases:

1. S-60 - Rootclaim’s main COVID-19 origins analysis — Rootclaim’s flagship Bayesian write-up: ~89% lab escape from gain-of-function research. The object of the $100k Challenge and Saar Wilf’s entry among the “six.”
2. S-73 - Peter Miller’s written case against the lab-leak theory — Miller’s central pre-debate written case for zoonosis at the Huanan Seafood Market; the position he defended (and won $100k defending) against Wilf, and his own entry among the “six” (given, by his own account, “half-jokingly”).

The debate record and its judges’ decisions (three more of the “six”): 3. S-70 - Judge Eric Stansifer’s written decision — the more technically extensive decision (~50pp incl. a Bayesian-computation tutorial); P(LL|evidence) ≈ 0.075% (~1-in-1300) by his own worked table, with explicit caveats about over-precision; also contains a pointed methodological critique of Rootclaim’s Bayes-factor construction. 4. S-67 - Judge Will van Treuren’s written decision — the other judge’s decision, modeled explicitly on Michael Weissman’s framework; final ratio ≈ 1-in-300 lab leak. 5. S-76 - The Rootclaim vs. Peter Miller COVID-origins debate, video record — the 3-part, ~18-20 hour video record itself (YouTube, 3 URLs on one node) plus the ~750 written slides referenced by both judges; the shared base material every other node here reacts to.

Post-debate synthesis: resolving the “six analyses spanning 23 orders of magnitude” claim: 6. S-79 - Scott Alexander’s ACX review of the Rootclaim debate — the primary source of the case’s framing claim; his comparison table is the “six” (Miller, Wilf/Rootclaim, judges Will and Eric, Alexander himself, and Daniel Filan), and he explicitly excludes Weissman from it. Also reports his own ~90-10-zoonosis estimate and a Good Judgment Project sub-study (~75-25 zoonosis). 7. S-63 - Rootclaim’s response to Scott Alexander’s ACX review — Rootclaim’s fullest unpacking of the “six estimates span 23 OOM” line (removing Miller: ~7 OOM; removing Wilf too: factor of ~50), argued as a HSM Bayes-factor dispute (Rootclaim: ~2x; most raters: up to 10,000x). 8. S-80 - Daniel Filan’s Bayesian analysis of COVID origins — the sixth analysis: an AI-alignment researcher’s same-day informal write-up, final posterior ≈ 1:24 (≈96% zoonosis), unusually explicit about its own low confidence.

Independent Bayesian analyses outside the “six,” and methodological cross-examination: 9. S-64 - Weissman’s ‘An Inconvenient Probability’ Bayesian analysis — a physicist’s independently-run, continuously-revised analysis (4:1-12:1 lab-leak-favoring in recent versions), explicitly named in the case brief but excluded from Scott’s six-way table. 10. S-83 - Judge Eric Stansifer’s response to Michael Weissman — Stansifer’s direct, itemized post-decision rebuttal of Weissman’s Bayes factors (concedes one point-vs-p-value conflation, disputes the WIV-engineering-attempt probability and the HSM-data-integrity concern). 11. S-86 - Levin’s NBER Bayesian assessment of COVID-19 origins — an economist’s Jan-2025 (post-debate, independent of it) Bayesian analysis using only spatiotemporal/zoonotic data, reaching 14,900:1 for lab leak — the opposite conclusion from most of the “six,” via a similar factorization method on overlapping data. Evidence the “conversation… continues to evolve” past the debate itself.

Ordering within each group is best-first by my rough quality×usefulness read; groups 1-2 are ordered debate-structurally (both sides, then the judges/record) rather than strictly by quality, since all five are close to equally load-bearing.

search_scope

Started from every URL given directly in the brief and in the shared search-plan.md, then: (1) fetched Scott Alexander’s ACX piece and Rootclaim’s response-to-Scott post as raw HTML via curl + a hand-written tag-stripper (not just WebFetch’s small-model summarizer, which on a first pass mis-attributed the “six estimates” table’s contents) to get verbatim text around “orders of magnitude” and resolve exactly who the six are: Scott’s own footnote reads “Peter, Saar, and the two judges all did their own Bayesian analysis. I followed along at home and tried the same. Daniel Filan… did one too” — confirmed by cross-reading Rootclaim’s response, which independently states “5 are by people who have never done a full probabilistic inference analysis… and one is by a team doing it for a decade” (matching Rootclaim as the lone professional entrant). (2) Extracted every <a href> around the “Weissman”/“Judge Eric”/“Filan”/“calculator” mentions in the raw ACX HTML to recover URLs stripped by prose-only summarization (this is how I found Filan’s Google Doc, Eric’s ermsta.com site incl. his Weissman response, and Scott’s blank calculator template). (3) Resolved the brief’s tinyurl redirects (via curl -L) confirming they point at exactly the two Google Drive files named in the brief, and downloaded + pdftotext’d both judges’ PDFs directly (Google Drive’s viewer doesn’t render for WebFetch, but direct PDF download + local text extraction worked cleanly and let me quote each judge’s actual final Bayes tables). (4) Fetched YouTube’s raw page HTML for all three debate-video IDs to get exact titles/upload dates/view counts rather than guessing. (5) WebSearched for the “other” candidate Bayesian analyses named in the search-plan (Quay & Muller, Demaneuf & de Maistre) plus new candidates (Daniel Filan, Andrew Levin/NBER) to check whether they belonged in the debate’s own “six” or were a separate, wider genre — resolved that Quay/Muller and Demaneuf are NOT part of the debate’s six and are not engaged by any of the six either; Levin’s NBER paper is a genuinely independent, post-debate entry to the wider genre and I minted it for that reason. (6) Downloaded and grepped the Rootclaim.com main analysis page and the Rootclaim blog’s “debate results” post directly for judge bios, tinyurls, and Yuri Deigin’s role.

exclusions

1. Rootclaim’s “Rootclaim’s COVID-19 Origins debate results” post (blog.rootclaim.com/rootclaims-covid-19-origins-debate-results/, 2024-02-18) — read in full as the hub linking both judges’ decisions/spreadsheets/video summaries and containing Rootclaim’s own “strawmanning”/Bayes-factor methodology explanation; not minted as its own node given the 11-node budget (its distinct content beyond what’s in S-63 - Rootclaim’s response to Scott Alexander’s ACX review is mostly about debate format, not new evidence/argument). Referenced in S-76 - The Rootclaim vs. Peter Miller COVID-origins debate, video record’s motivatedness field instead.
2. Judge Will’s probability spreadsheet and the video summaries of each judge’s decision (linked from the “debate results” post) and Scott Alexander’s blank Bayesian-calculator template — raw computational/video artifacts underlying already-minted prose decisions; not separately minted given budget.
3. Peter Daszak’s “half-jokingly” extreme estimate and the Good Judgment Project’s ~75-25 sub-study, both mentioned inside Scott Alexander’s ACX piece — informal asides without their own independent methodology write-up; covered in S-79 - Scott Alexander’s ACX review of the Rootclaim debate’s body rather than minted separately.
4. Quay & Muller’s Congressional-testimony-style Bayesian analysis, and Demaneuf & de Maistre’s “Outlines of a probabilistic evaluation of possible SARS-CoV-2 origins” (both candidates named in the shared search-plan as possibly part of “the six”) — checked and excluded: neither is part of the debate’s own six-way comparison and neither is engaged by any of the six analysts or the judges. They belong to a wider genre of independent COVID-origins Bayesian analyses that this slice’s mandate (the debate’s own record) doesn’t require covering exhaustively; flagged as a possible gap below rather than minted, since including them would have meant cutting a node that IS part of the debate’s direct record.
5. Andrew Gelman’s Columbia stats-blog commentary on the Bayesian-analysis genre (including a critique of Levin’s paper) — this is commentary/review of other people’s primary analyses, not itself a primary record of an argument about origins; excluded per the normal “no reviews as nodes” rule, which this slice’s exception does not extend to third-party commentary on others’ work.
6. Yuri Deigin (Rootclaim’s second-session debater, genetics) — his specific technical claims are genome-structure/engineering-signal arguments belonging to slice D; noted as a participant in S-76 - The Rootclaim vs. Peter Miller COVID-origins debate, video record’s authors field but not separately searched or minted here.

Slice shape

A tight, self-referential 11-node web rather than 11 independent finds: nearly every node here explicitly reacts to 1-3 others in the same set (Rootclaim ↔ Scott Alexander ↔ the judges ↔ Weissman ↔ Filan), which is the point — this slice’s job was to pin down exactly who the case’s “six independent analyses spanning 23 orders of magnitude” are (resolved with high confidence: Peter Miller, Saar Wilf/Rootclaim, judges Will van Treuren and Eric Stansifer, Scott Alexander, and Daniel Filan — NOT Weissman, who Scott explicitly excluded from his table, and NOT the various other named Bayesian COVID-origins analyses floated as candidates in the shared search-plan) and then to add the two most significant pieces of Bayesian reasoning about this evidence that sit just outside that six: Weissman’s analysis (explicitly named in the brief) and its direct rebuttal by judge Eric, plus one genuinely independent post-debate replication (Levin’s NBER paper, Jan 2025) that reaches the opposite conclusion from most of the six using a similar method — good evidence for the case framing’s claim that “this intense epistemic effort represents a point in time in a conversation which continues to evolve.” I went looking for a standalone empirical primary establishing the base rate of “novel pathogens’ first-detected clusters centering on a market/venue regardless of true origin” — an assumption several of these analyses (Rootclaim’s response, Weissman, Eric) lean on heavily and dispute the size of — and could not find one; if no sibling slice (A/B/F) has separately minted such a study, that looks like a real empirical gap rather than a search failure on my part.