Bias Audit — Wave 1 (2026-06-11)

Scope: all 51 content notes in 01_prehistoric … 09_comparative (37 tier-1 sources, 13 tier-2 notes, 1 tier-3 synthesis), plus 00_meta/OPEN_QUESTIONS.md, audited against AGENTS.md, 00_meta/Methodology.md, and all six templates. Method: full read of every file, plus systematic rg scans for frontmatter completeness, enum validity, url/url_verified consistency, confidence×evidence-class pairing, and wiki-link resolution.

Headline: no missing mandatory frontmatter field anywhere (a genuinely clean scan — every claim has counter_evidence, every motif has a transmission enum, every source has url_verified). The violations live one level down: in how fields are used, in cross-domain contradictions, and in the workflow loop.

The field-presence scan passed 51/51. The violations are semantic:

Broken wiki-links: 28 distinct targets do not exist. Worst offenders: every Feeds into section written by the 02, 03, 05, 07 gatherers names tier-2 notes that were never written (el-yahweh-relationship, canaanite-substrate-claim, pie-dragon-slaying-formula ×3, shinto-origins-archaeology-vs-text, ubaid-religious-continuity, etc.). Two are wrong slugs for notes that do exist: dying-and-rising-god → actual dying-rising-god (in 05/israelite-religion-origins.md) and indo-european-sky-father → actual sky-father (in 06/vedic-religion.md). Full list preserved in the scan; the two wrong slugs are the cheapest fixes with the highest navigation payoff.

Orphaned source: 07_east_asian/1_sources/hardacre-shinto-history-diversity.md feeds into two notes that don't exist and is cited by nothing. A tier-1 note with zero tier-2 consumers and phantom feeds-into links is gathered-but-unused inventory.

• 05_abrahamic — school over-weight, the clearest case. Cross (Harvard) + Smith (Cross's student, same framework — the note's own reliability section admits it) + Blenkinsopp (endorses the Kenite hypothesis Cross also accepts). All three tier-1 sources sit inside one broadly continuist consensus. The dissenters (Day, Emerton, Hadley, Kaufmann, Tebes) appear only inside counter-evidence prose, never as tier-1 records. The steelman is being written from memory of the opposition, not from gathered opposition.
• 04_indo_european — all three sources are pro-reconstruction comparativists (Mallory-Adams, Watkins, West). No skeptical source on the limits of comparative mythology. Counter-evidence is again internal.
• *08_indigenous — all three sources are proponents of bold contested theses* (Lewis-Williams's shamanism model, Nunn-Reid's deep-time oral memory, Witzel's Laurasian/Gondwana scheme). The notes themselves calibrate beautifully (confidence: low, twice), but the critics doing the real work in those notes — Henige, Hiscock, Bahn, Hutton — have no tier-1 records. The domain gathered the claims and cited the critiques from memory.
• 06_dharmic and 07_east_asian — the models to copy. 06 pairs a revisionist (Bronkhorst) against the mainstream (Olivelle) plus a deliberately-included methodologically-weak historical source (Marshall 1931) used as an object of critique. 07 pairs Keightley vs Eno as named interpretive rivals on Di. This is what per-domain diversity should look like.
• 09_comparative — no methodology over-weight found. Smith vs Mettinger is an explicit steelman pair; philology (West), epigraphy (Linear B), and areal statistics (Berezkin) are distinct methods. The one gap: function-theory scholarship in the synthesis (Durkheim, Norenzayan, Sosis, Bentzen, Malinowski, the Whitehouse retraction) has zero tier-1 records — already self-filed as Q2/Q8/Q9, so noted, not double-counted.
• Era skew (vault-wide): 02's textual anchors are 1969/1981 editions accessed through tertiary summaries; nothing in 02 post-dates Lambert 2013, and Wasserman 2020 is mentioned but not used.

The hard rule — high resting on class 3-reconstruction/4-ethnography — is violated zero times in frontmatter. Genuinely impressive: 04 keeps the PIE deity claim at medium with attestation_earliest: speculative; 08 holds both deep-time claims at low; 09's high verdicts are scoped and class-2 anchored.

Real calibration problems found instead:

1. 02/atrahasis-antedates-gilgamesh-flood-tablet.md — confidence: high violates the AGENTS §2.4 Wikipedia cap. Its sole tier-1 source is the atrahasis source note whose extractions are explicitly grounded in "Wikipedia: Gilgamesh flood myth" and a COJS summary. AGENTS.md: tertiary-grounded claims are "capped at medium until upgraded." The claim is almost certainly correct (09's independently-gathered George 2003 note supports it), but as cited within its own domain it is high-on-Wikipedia. Either re-source 02 against Lambert & Millard / George directly, or drop to medium.

2. The same descent verdict carries medium in 04 and high in 09. 04/dyeus-sky-father-cognate-set.md concludes the claim "cannot be elevated to high without either Anatolian evidence or direct PIE texts"; 09/sky-father.md rates the descent verdict high, citing Anatolian evidence (šiu-, Luwian Tiwaz) that 04 says doesn't exist (see §6). The scoping differs slightly (deity-reconstruction vs name-formula descent), but no reader can recover that from the frontmatter, and neither note acknowledges the other.

3. 05/el-yahweh-identification-canaanite.md — body says "high for the presence of El-identification, medium for the mechanism"; frontmatter says only high. The split is the honest value; the single field overstates.

4. Single-source highs: 02/atrahasis-antedates… (1 source), 03/pyramid-texts-oldest-large-corpus.md (1 source, and a world-superlative claim whose own counter_evidence concedes the Sumerian comparison is unresolved). Both satisfy the ≥1-source letter; neither satisfies its spirit at high.

5. Speculative-labeling is otherwise good — 04's attestation_earliest: speculative, 08's split anchors ("transmission chain undated"), and 09's "(composition inferential)" tags are exactly per Methodology. The single mislabel is 06/vedic-religion.md (§1).

Sky-father (04 vs 09) — contradictions, not just divergence:

The Anatolian rows are genuinely irreconcilable as written: West 2007 (cited by both domains) treats šiu-/Tiwaz as root reflexes with the caution that the figure must be argued — 09 transmits the caution, 04 flatly denies the cognates. One of them must change.

Rigveda manuscript date — three-way contradiction: 06/rigveda-oral-composition-attestation-gap.md says oldest ms. c. 1040 CE, Nepal; 04/rigveda-oral-gap-composition-vs-attestation.md says 1464 CE, BORI Pune ("No earlier physical manuscript … is known to exist in any collection" — directly contradicted by 06); 09/rigveda-dyaus.md says c. 14th c. CE. These are two near-duplicate claim notes (04 and 06 cover the same atomic claim — itself a one-claim-per-note breach across domains) plus a source note, none linking the others. The oral-gap arithmetic differs by ~400 years depending on which you read.

Flood (02 vs 09) — broadly consistent, three minor divergences: (a) SB Gilgamesh redaction dated "c. 1300–1000 BCE" in 02 vs "c. 1200–1100 BCE" in 09; (b) 02's counter-evidence holds open that OB Gilgamesh "may not have contained" a flood account where 09 (per George) states the OB version "did not include the flood story" — strength mismatch; (c) 02 hedges that Eridu Genesis "may represent an … older tradition (possibly oral c. 2800 BCE)" while 09 leans on Civil's suspicion that the Sumerian text is dependent on the Akkadian — opposite leans, both flagged as open, acceptable but worth one reconciling sentence. Verdict agreement is solid: both call Genesis←Mesopotamia contact. One enum misuse: 02 labels the Gilgamesh-XI-from-Atrahasis relation "descent" — by the vault's definitions (descent = inherited from common ancestor) direct literary incorporation within one scribal culture is not descent; 09 correctly describes it as textual dependence without forcing the enum.

Attestation dates that do agree (recorded for fairness): Atrahasis tablets (Ammi-ṣaduqa reign — "1646–1626" vs "c. 1635" are the same epigraphic fact), Eridu Genesis tablet c. 1600 BCE, Pyramid Texts c. 2350 BCE (03 ↔ 09), Kuntillet Ajrud c. 800 BCE, oracle bones 1254–1197 BCE.

• No invented-looking URLs. All nine recorded URLs are plausibly real (DOI paths match the cited articles for Hovers 2003, Blenkinsopp 2008, Nunn-Reid 2016; the escholarship ID matches Stevenson's known entry).
• The three url_verified: yes (all in 02) are honest about access but dishonest in role: what was verified is a Wikipedia/worldhistory page, recorded in the url field of a note whose citation is a print monograph. The result is a note that looks like a verified scholarly source and is a verified tertiary summary. This is the single most corrosive pattern found, because every downstream confidence judgment inherits it silently (§4.1).
• Four not-online with URLs present (§1) — wrong enum, low malice: three are paywalled DOI pages, but Stevenson's is an open-access PDF that could simply have been fetched and marked yes.
• 04's two url_verified: no entries are the correct honest pattern for known-but-unfetched URLs.

Counts: 4 critical · 6 major · 10 minor.

1. Fix the 02 source chain first (#1–#2). It is the only place where the vault's strongest stated taboo-adjacent pattern (tertiary grounding presented as scholarly grounding) feeds a high confidence. The repair is cheap because 09 already gathered the proper editions for the same texts — 02 can largely re-cite them.

2. Run a reconciliation pass before any new gathering on overlapping topics. Rigveda manuscripts (#3) and Anatolian sky-father evidence (#5–#6) prove that two agents writing about the same fact will diverge. Concretely: any wave-2 note touching a topic with an existing note in another domain must link it and state agreement/disagreement in one sentence.

3. Restore the Open Questions loop (#4): harvest the ~30 stranded domain-note questions into the register, then have wave-2 gatherers draw assignments from it — the register is currently synthesizing-agent monoculture (Q1–Q11 all from 00/09), which is itself a bias channel.

4. Buy one dissenting source per skewed domain (#10). The counter-evidence fields are currently written from the gatherers' memory of critics; gathering Day 2000, Henige's response, and a comparative-mythology critique converts remembered steelmen into sourced ones — and tests whether they survive contact with the actual texts.

5. Make two scans part of the wave gate: (a) wiki-link resolution (this wave: 28 broken), (b) url ↔ url_verified consistency (this wave: 7 files). Both are mechanically checkable; neither should reach the human reviewer again.

6. Tradition-profile counter-evidence standard: require at least one counter-evidence prong per major section of a tradition profile (origins, continuity, functions), or explicitly scope the confidence field to the counter-evidenced part (§2 pattern).

7. Keep doing what 06/07/09 did: rival-source pairing (Keightley/Eno, Smith/Mettinger, Bronkhorst/Olivelle) produced the vault's best-calibrated notes. That design choice — not reviewer vigilance — is the scalable bias control.