Step 1 — Search plan (planner 1a), COVID-19 origins

Main question (verbatim): “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)?” curated_target_N = 5. Step-1 total budget for this run (orchestrator’s lean override for the test, not the mode file’s default 8N/4N formula): ~12 source nodes written, ~16 read properly, split into exactly 2 slices (~6 written / ~8 read each), plus unbounded skimming of reviews/discovery hubs on top of that.

Survey: literatures, sides, recurring datasets

The debate has two clean evidence families, and the run’s own scope notes already carve it this way (“geographic/temporal signal of the earliest cases” + “market-vs-lab ascertainment bias” vs. “genomic features (furin cleavage site, RaTG13/BANAL relatedness, restriction-site/lineage-A-vs-B patterns)”):

Epidemiological / field evidence — who got infected when and where, what was being sold at the Huanan market, what its environmental swabs contained, and what early-case/sampling data China released vs. withheld. Zoonosis-favoring anchors: Worobey et al. 2022 Science (market-centered case geolocation/spatial clustering); the China CDC-led Huanan-market environmental/animal-swab surveillance paper and the international team’s reanalysis of that same raw data (susceptible-species DNA co-located with virus-positive swabs); Xiao et al. 2021 (wildlife species actually sold at Wuhan markets pre-pandemic). Lab-leak-favoring/critical anchors: Bloom 2021 (recovered deleted early-Wuhan SRA sequences, an ascertainment-bias critique of the market-centric case series); critiques of Worobey 2022’s dataset/conclusions found while snowballing.
Genomic + institutional/biosafety evidence — features of the virus’s own sequence, its closest known relatives, and the paper trail at the institutions implicated in the “research-related incident” hypothesis. Zoonosis-favoring anchors: Andersen, Rambaut, Lipkin, Holmes & Garry 2020 (“Proximal Origin”, furin cleavage site explicable by natural evolution); Temmam et al. 2022 (BANAL bat-CoV genomes from Laos, closer than RaTG13 at some genome regions); Pekar, Worobey, Wertheim et al. 2022 (molecular-clock case for two independent spillovers, lineage A vs. B). Lab-leak-favoring anchors: Segreto & Deigin 2021 (furin-cleavage-site engineering-signature argument); the restriction-site/“synthetic fingerprint” claim (Bruttel, Washburne & VanDongen — peer-review status unverified, flag it); the leaked DEFUSE grant proposal (2018 DARPA submission proposing furin-site insertion experiments in SARS-related bat CoVs); the WIV virus/sample database taken offline Sept. 2019; US government assessments (ODNI/DOE/FBI, split across agencies) and the House Select Subcommittee’s 2024 report.

Neither slice is “pro-zoonosis” or “pro-lab-leak” as a whole — each carries both sides on its own axis. That is what makes the split by data axis rather than by side: cutting by side would leave each slice internally balanced on data type but the two slices duplicating each other’s re-analyses; cutting by axis (as done below) makes the two slices rest on genuinely different underlying data-bases, which is the independence the task asks for.

Recurring datasets to flag for step 2’s shared D-nodes (do not act on this yourselves at step 1 — no data_basis field here, just say so in prose in summary/relevance_note):

The Huanan market environmental-swab dataset (China CDC’s own sampling, later partly released on GISAID): the China-CDC paper and the international reanalysis are two independent interpretations of the same raw swabs — both worth a node, but a third re-analysis of those same swabs would not be.
Worobey 2022’s compiled early-case geolocation record (assembled from leaked/court-obtained/WHO-report fragments) is itself a single dataset — later citations of its numbers are not independent new evidence.
The SARS-CoV-2 reference genome / GISAID sequence collection underlies essentially every genomic argument (furin site, restriction sites, lineage A/B molecular clock). That’s expected and fine — each paper targets a different genomic feature with a different method, not a restatement — but still worth flagging in prose.
RaTG13’s own sequencing data/metadata (Zhou et al. 2020, WIV) and the BANAL genomes (Temmam et al. 2022, separate Laos expedition) are two distinct “closest relative” datasets — don’t let them get conflated into one.

Why none of the 8 seed items get their own source node. The seed material — Scott Alexander’s ACX writeup, Judge Will’s + Judge Eric’s decisions, Michael Weissman’s analysis, Rootclaim’s response, and the 3 debate videos — is commentary/synthesis over the same underlying evidence; none of it “made the measurement, ran the experiment, or holds the record” for the object-level question. Per the mode file’s primary-sources-only rule, all 8 are discovery hubs: mine them for the primary papers/datasets/documents they cite or discuss, and node those. This operationalizes the orchestrator brief’s “step 1 MUST reach the primary artifacts behind the seed summaries, not stop at the summaries.”

Slice division & rationale

Exactly 2 slices, split by data axis, each carrying both zoonosis- and lab-leak-favoring sources on its own axis:

slice	owns	written	read
1 — epi-field-evidence	case geography/timing, market wildlife-trade & environmental samples, data-suppression/ascertainment-bias record	6	8
2 — genomic-institutional-evidence	furin cleavage site, closest viral relatives (RaTG13/BANAL), lineage A/B molecular evolution, WIV documents & database, government/congressional investigations	6	8
total		12	16

Boundary rule of thumb: evidence about who/where/when got infected or what physical animals/samples were at the market → slice 1; evidence about the virus’s own sequence or the research institutions’ documents/practices → slice 2. Named boundary call: Pekar/Worobey/Wertheim 2022 combines case-timing correlation with sequence-divergence analysis, but its core method is phylogenetic/molecular-clock and the scope notes group “lineage-A-vs-B patterns” under genomic features — it’s slice 2’s; slice 1 should not also node it despite Worobey’s overlapping authorship with the slice-1 market paper.

Collision avoidance (the two searchers run in parallel with no live coordination): each prompt below names the counterpart’s candidate items explicitly. If a searcher lands on an item that’s clearly the other slice’s, it should skip it and log it under its own exclusions as “out of scope — slice X,” not mint a duplicate node.

Gaps anticipated

The judge decisions are Google-Drive-hosted PDFs — may be gated/flaky. Fallback: the ACX post quotes both judges extensively; use that if the PDFs don’t load, and note the gap.
The debate videos run ~15h total across 3 videos — searchers should not transcribe/watch in full. The text seed docs (ACX post, judge decisions, Weissman, Rootclaim response) cover the same ground far more cheaply; videos are only for spot-checking a specific claim not found in text.
Given probable deliberate data suppression (scope note 2), some “primary artifacts” on the epi side may not exist in accessible form at all (e.g. the full original early-case line list may never have been released). That is an expected, documentable finding, not a search failure — node the best available near-primary (e.g. reporting establishing what was/wasn’t released) and say so plainly.
Some lab-leak-side genomic arguments (e.g. the restriction-site claim) may be preprint-only or have contested peer-review status. Include them anyway — motivatedness/publication status is data step 2 weighs later, not a filter now.

Balance/independence self-check

Both slices carry named zoonosis- and lab-leak-favoring anchors (not one slice per side). Yes.
Slices split on independent data-bases (epi/field records vs. genome sequences + institutional documents), not on re-analyses of one dataset. Yes.
Shared/reused datasets (market swabs, case-geolocation compile, GISAID genome collection, RaTG13-vs-BANAL) flagged explicitly so step 2 can build shared D nodes and avoid a curated set that is secretly one dataset. Yes.
Seed summaries (ACX/judges/Weissman/Rootclaim/videos) used only as discovery hubs, never noded — forces the search past them to primaries. Yes.
6 written / 8 read per slice = 12/16 total, matching the orchestrator’s lean budget exactly. Yes.

SEARCHER PROMPTS (spawn each as-is; both already carry the TEST-RUN NOTICE)

Searcher prompt 1b-A — slice: epi-field-evidence

Read .claude/skills/flf-epistack/steps/step-01-find-sources.md — it is your full instruction set. You are a 1b searcher. Analysis directory: projects/create a useful aligning (AI-)macroagents agenda/Participate in FLF competition/analysis-tests/covid1 (Quote the path in shell commands — it has spaces and parentheses; cwd = vault root /home/simonskade/workspace.) Main question (verbatim): “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)?”

Your slice: epi-field-evidence — the epidemiological/field-evidence axis: (i) geographic and temporal signal of the earliest human cases, (ii) the Huanan Seafood Market’s wildlife trade and environmental (swab) samples, (iii) market-vs-lab ascertainment-bias arguments, and (iv) the record of which early-case/sampling data China released vs. withheld. Boundary rule: if the evidence is about who/where/when got infected or what physical animals/samples were at the market, it’s yours. You must NOT touch slice 2 (genomic-institutional-evidence) — the virus’s own sequence (furin cleavage site, RaTG13/BANAL relatedness, lineage-A/B molecular-clock argument) and the research-institution documentary record (DEFUSE grant, WIV database, government/congressional investigations). If you land on one of those, skip it and log it in your exclusions as “out of scope — slice 2”; do not mint it.

Balance — represent both directions. Zoonosis-favoring anchors on your axis: Worobey et al. 2022, Science (“The Huanan Seafood Wholesale Market… was the early epicenter of the SARS-CoV-2 pandemic” or similar title — verify exact title/DOI; case geolocation/spatial clustering); the China CDC-led Huanan-market environmental/animal-swab surveillance paper (Liu et al., eventually Nature 2023 after a long preprint period — verify authors/venue/date); the international reanalysis of that same raw swab data reporting susceptible-species (e.g. raccoon dog) DNA co-located with virus-positive samples (Crits-Christoph, Débarre, Worobey et al., first posted ~March 2023 — verify final venue); Xiao et al. 2021, Scientific Reports (wildlife species actually sold at Wuhan markets 2017–2019). Lab-leak-favoring/critical anchors: Bloom 2021 (recovered deleted early-Wuhan SRA sequences — verify exact venue; argues market cases may not be the true earliest/index cases, i.e. an ascertainment-bias critique); any rebuttal/critique of Worobey 2022’s dataset or conclusions you find while snowballing. Neutral/contested: the WHO-convened Global Study of Origins of SARS-CoV-2, China Part (WHO-China Joint Report, March 2021) — the original field mission’s epi-curve and market-investigation data; node it even though its “extremely unlikely [lab leak]” framing is contested, since it holds primary field data.

Also actively look for (per scope note 2 in initial_prompt.md): concrete reporting or documents establishing which early-case records/line lists/sampling data were released vs. withheld, and the market’s closure-and-disinfection timeline before international investigators arrived. This may surface news/investigative-journalism sources (source_type: news) rather than journal articles — that’s fine. Treat absence of evidence carefully: if a record was never released, say so plainly in the summary (a real, documentable finding) rather than silently dropping the topic or treating the absence itself as proof of either side.

Set motivatedness wherever there’s a known angle: Chinese state-affiliated authorship on market/case data (government control over what’s released); Worobey-team members’ publicly stated zoonosis-favoring priors; any source with a stated institutional stake.

Recurring-dataset awareness — say so in prose in summary/relevance_note; do NOT write a data_basis field, that’s step 2’s: the China-CDC swab paper and the Crits-Christoph/Débarre reanalysis share the same raw environmental samples — both are worth nodes (independent interpretations), but don’t add a third re-analysis of those same swabs. Worobey 2022’s case-geolocation dataset is itself a single compiled record — later papers merely re-citing its numbers aren’t independent new evidence.

Discovery hubs to mine for primaries — read these for citations, do NOT give any of them their own source node: Scott Alexander’s ACX writeup (https://www.astralcodexten.com/p/practically-a-book-review-rootclaim); Judge Will’s decision (https://drive.google.com/file/d/1YhmkYB32RpGsXvQTsX4xZ0Yul1wiwh8Z/view) and Judge Eric’s decision (https://drive.google.com/file/d/1aHlhPd-16EOabzXhiajT5PBm3uVCAG3T/view) — Google-Drive-hosted, may be flaky; if inaccessible, fall back on the ACX post’s extensive quotation of both and note the gap; Michael Weissman’s analysis (https://michaelweissman.substack.com/p/an-inconvenient-probability-v57); Rootclaim’s response (https://blog.rootclaim.com/covid-origins-debate-response-to-scott-alexander/); and the 3 debate videos (https://www.youtube.com/watch?v=Y1vaooTKHCM, https://www.youtube.com/watch?v=KdORmvU8MLI, https://www.youtube.com/watch?v=d1dbfoK8nSE — ~15h total; do NOT transcribe/watch in full, the text docs above cover the same ground faster, only spot-check video timestamps if chasing a claim not in the text). Pull only the epi/field-relevant citations from these; leave genomic/institutional citations for slice 2. If you find a general scientific review covering the epi evidence cheaply (e.g. Holmes et al. 2021, Cell, “The origins of SARS-CoV-2: A critical review” — verify), mine it too under the same rule: discovery hub, not a node.

Budget: write 6 source notes, read ~8 sources properly (open, extract real metadata, write a real summary) — skim further via the discovery hubs above at no cost to that count.

Mint every node through the script — never hand-pick an id: python3 .claude/skills/flf-epistack/scripts/create_node.py “projects/create a useful aligning (AI-)macroagents agenda/Participate in FLF competition/analysis-tests/covid1” —type source —title “Descriptive title, no id” (The --title value shown is the mode file’s own placeholder phrasing — replace it with each node’s actual descriptive title, with no id in it. Pipe the node’s markdown — frontmatter + body, with {{ID}} wherever the bare id belongs, at minimum id: {{ID}} — on stdin.) Do not spawn subagents.

Before returning, write your orientation note at agent-notes/orientation/epi-field-evidence.md per the mode file’s spec: your sources as [[wikilinks]] grouped under topic headings (best-first within each group), search_scope, exclusions (including anything you skipped as out-of-scope for slice 2), and ~1 paragraph on your slice’s shape and any gaps.

Return to your orchestrator: the S-ids you created, one line on any gaps you hit, and a short 1-line-per-source list.

┌─ TEST-RUN NOTICE — copy verbatim into every subagent brief ─────────────┐

This is a quick test run under time pressure, not a production analysis. The goal is only to surface major errors or problems in the pipeline, not to produce an optimal output. Work fast, and don’t worry about your output being imperfect or sub-optimal — a rough, correct-enough pass is exactly what’s wanted; do not polish or iterate. Do not log minor imperfections. If you hit a real problem — a pipeline bug, a broken or ambiguous cross-step interface, missing or contradictory instructions, a script that errors, or anything that blocks or corrupts the run — append ONE entry to the problem log with a severity tag (BLOCKER or MAJOR) and one line of context (which step, which slice, what happened). The log is at: projects/create a useful aligning (AI-)macroagents agenda/Participate in FLF competition/analysis-tests/covid1/problem-log.md Append with a single write (e.g. a >> shell append) so parallel appends don’t clobber each other; never edit or delete other agents’ entries. Then keep going if you can.

└──────────────────────────────────────────────────────────────────────────┘

Searcher prompt 1b-B — slice: genomic-institutional-evidence

Your slice: genomic-institutional-evidence — the virus’s-own-sequence axis plus the research-institution documentary record: (i) the furin cleavage site, (ii) closest known viral relatives (RaTG13, BANAL genomes) and what they imply about divergence time/recombination, (iii) the lineage-A-vs-B / restriction-site molecular-evolution argument, and (iv) WIV research documents/database and government/congressional investigations into a possible research-related incident. Boundary rule: if the evidence is about the virus’s own genome or the research institutions’ documents/practices, it’s yours. You must NOT touch slice 1 (epi-field-evidence) — case geolocation/timing, market wildlife-trade records, market environmental swabs, and data-suppression/ascertainment-bias reporting. If you land on one of those, skip it and log it in your exclusions as “out of scope — slice 1”; do not mint it.

Balance — represent both directions. Zoonosis-favoring anchors: Andersen, Rambaut, Lipkin, Holmes & Garry 2020, Nature Medicine, “The Proximal Origin of SARS-CoV-2” (furin cleavage site explicable by natural evolution); Temmam et al. 2022, Nature, “Bat coronaviruses related to SARS-CoV-2 and infectious for human cells” (BANAL-52/-103/-236 genomes from Laos, closer than RaTG13 at some genome regions); Pekar, Worobey, Wertheim et al. 2022, Science, “The molecular epidemiology of multiple zoonotic origins of SARS-CoV-2” (molecular-clock case for two independent spillovers, lineage A vs. B — this belongs here, not slice 1, despite Worobey’s overlap with slice 1’s market paper: it’s a genomic/sequence-divergence analysis, a different primary artifact and dataset). Lab-leak-favoring anchors: Segreto & Deigin 2021, Bioessays (furin-cleavage-site engineering-signature argument); the restriction-site/“synthetic fingerprint” claim (Bruttel, Washburne & VanDongen — verify exact title/venue/peer-review status, may be preprint-only or contested); the leaked DEFUSE grant proposal (EcoHealth Alliance/WIV/UNC submission to DARPA’s PREEMPT program, 2018, rejected, leaked 2021 via DRASTIC/The Intercept — proposed inserting furin cleavage sites into SARS-related bat coronaviruses); reporting on the WIV virus/sample database taken offline September 2019. Neutral/mixed: Zhou et al. 2020, Nature (first report of RaTG13 — baseline relatedness data, cited by both sides); US government assessments (ODNI unclassified reports 2021 + later declassified material, DOE “low confidence” and FBI “moderate confidence” lab-leak assessments as reported ~2023, other agencies’ zoonosis-leaning low-confidence assessments) — capture the primary government document itself where you can, not just news paraphrase; US House Select Subcommittee on the Coronavirus Pandemic final report, Dec. 2024 (primary document holding subpoenaed EcoHealth/NIH/WIV-related communications, lab-leak-leaning conclusion — note the subcommittee’s Republican-majority composition as motivatedness).

Set motivatedness wherever there’s a known angle: EcoHealth Alliance/Daszak-authored or -linked items (professional/reputational stake in the zoonotic narrative, given DEFUSE and prior WIV funding ties); DRASTIC-linked material (an explicitly lab-leak-hypothesis-driven citizen-investigator group); the congressional report’s political composition; Chinese state control over what WIV database contents were ever released.

Recurring-dataset awareness — say so in prose in summary/relevance_note; do NOT write a data_basis field, that’s step 2’s: nearly every genomic argument here re-reads the same SARS-CoV-2 reference genome/GISAID sequence collection — expected and fine, since each paper targets a different genomic feature (furin site vs. restriction sites vs. phylogenetic distance) with a different method, not a restatement — but still flag it. Keep RaTG13’s own sequencing data/metadata (Zhou et al. 2020, WIV) and the BANAL genomes (Temmam et al. 2022, separate Laos expedition/institution) as two distinct data-bases — don’t conflate the two “closest relative” datasets.

Absence-of-evidence caveat (scope note 2 in initial_prompt.md): the WIV database’s Sept-2019 takedown and China’s non-cooperation with follow-up investigation requests are suspicious circumstantially but are absence of evidence, not evidence of a specific mechanism — node what’s actually documented (e.g. reporting establishing the takedown date and what has/hasn’t been recovered since) rather than overclaiming in your summary.

Discovery hubs to mine for primaries — read these for citations, do NOT give any of them their own source node: Scott Alexander’s ACX writeup (https://www.astralcodexten.com/p/practically-a-book-review-rootclaim); Judge Will’s decision (https://drive.google.com/file/d/1YhmkYB32RpGsXvQTsX4xZ0Yul1wiwh8Z/view) and Judge Eric’s decision (https://drive.google.com/file/d/1aHlhPd-16EOabzXhiajT5PBm3uVCAG3T/view) — Google-Drive-hosted, may be flaky; if inaccessible, fall back on the ACX post’s extensive quotation of both and note the gap; Michael Weissman’s analysis (https://michaelweissman.substack.com/p/an-inconvenient-probability-v57); Rootclaim’s response (https://blog.rootclaim.com/covid-origins-debate-response-to-scott-alexander/); and the 3 debate videos (https://www.youtube.com/watch?v=Y1vaooTKHCM, https://www.youtube.com/watch?v=KdORmvU8MLI, https://www.youtube.com/watch?v=d1dbfoK8nSE — ~15h total; do NOT transcribe/watch in full, the text docs above cover the same ground faster, only spot-check video timestamps if chasing a claim not in the text). Pull only the genomic/institutional-relevant citations from these; leave epi/field citations for slice 1. Also mine, under the same rule (discovery hub, not a node): Holmes et al. 2021, Cell, “The origins of SARS-CoV-2: A critical review” (verify — broad multi-author review; take its genomic-argument citations); optionally Alina Chan & Matt Ridley’s book “Viral: The Search for the Origin of Covid-19” (2021) as a lab-leak-leaning synthesis — treat as a discovery hub too unless it turns out to contain genuinely original primary reporting (documents the authors themselves obtained), in which case that specific item can be its own source.

Budget: write 6 source notes, read ~8 sources properly (open, extract real metadata, write a real summary) — skim further via the discovery hubs above at no cost to that count.

Before returning, write your orientation note at agent-notes/orientation/genomic-institutional-evidence.md per the mode file’s spec: your sources as [[wikilinks]] grouped under topic headings (best-first within each group), search_scope, exclusions (including anything you skipped as out-of-scope for slice 1), and ~1 paragraph on your slice’s shape and any gaps.

Return to your orchestrator: the S-ids you created, one line on any gaps you hit, and a short 1-line-per-source list.

┌─ TEST-RUN NOTICE — copy verbatim into every subagent brief ─────────────┐

EpiStack

Explorer

search-plan