Audit dashboard

This dashboard is auto-regenerated on every build from the wiki’s audit harness. It is the single transparency surface for reading “what is the current Cochrane-level defensibility of the wiki, by dimension.” External reviewers, regulatory affairs, and brand legal teams auditing the wiki should land here.

Generated: 2026-05-17. Baseline snapshot: 2026-05-17-baseline.

Scorecard

Dimension	Score	Status
Taxonomy v2.0 coverage (destination pages exist for HMTc rows)	277 / 277 (100%)	PASS
Routing audit (unresolved source routings)	0 unresolved	PASS
Routing rows generated	661	—
Ingredient mandatory sections (non-stub pages)	237 / 237 (100%)	PASS
Product mandatory sections (non-stub pages)	84 / 85 (99%)	WARN
Routing discipline check 1 (sources with product-shaped matrices but empty products[])	1 failures	WARN
Routing discipline check 2 (HMTc-locked products with zero routed sources)	0 of 290	INFO
Source page DOI completeness	0 missing	PASS
Claim provenance — body-prose numerical claims with traceable citation	3845/3845 (100%)	PASS

Taxonomy v2.0 coverage

Total subcategories defined by HMTc Taxonomy v2.0: 277. Pages scaffolded or content-filled: 277 (100%). Missing: 0.

• Exact frontmatter declaration (hmtc_category + hmtc_row): 260
• Slug-alias match (frontmatter needs backfill): 17
• Missing destination pages: 0

Detail report: taxonomy-coverage-audit.md. Full PRISMA-equivalent: coverage.

Routing audit

Routing layer (CLAUDE.md Part 5b) fans source-page declarations of products: [<slug>] to product-category pages. 661 routing rows in current audit.

• Unresolved targets (source frontmatter declares a slug that doesn’t exist): 0
• Malformed sources (missing required frontmatter fields): 568 (non-blocking)

Detail: product_source_routing_audit.csv, routing_unresolved.csv, routing_malformed.csv.

Routing discipline

Check 1 — sources with product-shaped matrices: slugs but empty products: [] and no pending_category_lock flag. These are sources that should route to a product page but the frontmatter doesn’t declare it.

• Sources scanned: 920
• Sources with product-shaped matrices: 358
• Failures: 1

Check 2 — HMTc-locked product pages (carrying hmtc_category + hmtc_row) with zero routed sources. These are scaffolded destinations that haven’t received their Phase 1 ingest yet.

• HMTc-locked products: 290
• Products with no routed sources: 0

Check 2 is informational, not a failure: the scaffolded universe (277 taxonomy rows) is larger than the current ingested universe (897 source pages); rows without sources are the priority queue for the next ingest pass, not defects.

Detail: routing-discipline-audit.csv.

Ingredient mandatory sections

Non-stub ingredient pages must carry 8 H2 sections per CLAUDE.md Part 6 (Why this commodity accumulates / Heavy metal contamination profile / Ranges by source, region, and variety / Processing effects / Ingredient-derivative risk / Mitigation options / Regulatory limits that apply / Sources). Stubs (untouched pages with all-pending contamination_profile) are exempt.

• Non-stub passing: 237 of 237 (100%)
• Non-stub failing: 0

Detail: ingredient-mandatory-sections-audit.csv.

Product mandatory sections

Non-stub product-category pages must carry 10 H2 sections per CLAUDE.md Part 6 (Who this page is for / Methodology / Literature Evidence Summary / Source Evidence Inventory / Broad Product Context / Federal/Regulatory Limits / Levers to reduce contamination / How standards math uses this page / Historical recalls and enforcement / Sources). Stubs (scaffolded pages with no routed sources) are exempt.

• Non-stub passing: 84 of 85 (99%)
• Non-stub failing: 1
• Stubs (exempt): 235

Detail: product-mandatory-sections-audit.csv.

Source page corpus

Total source pages in wiki/sources/: 920.

By evidence tier:

• A: 804
• B: 112
• C: 4

By source type:

• agency-report: 3
• book-chapter: 3
• conference-abstract: 1
• conference-proceedings: 1
• dataset: 1
• gov-data: 5
• gov-guidance: 3
• gov-regulation: 1
• gov-report: 53
• government-guidance: 1
• government-regulatory: 2
• government-report: 32
• government_dataset: 5
• government_documentation: 1
• government_report: 3
• industry: 1
• industry-application-note: 1
• news-article: 1
• news-commentary: 1
• ngo-report: 2
• nonprofit: 1
• peer-reviewed: 754
• regulation: 4
• regulatory: 2
• review: 31
• review-chapter: 1
• study: 1
• systematic-review: 1
• textbook-chapter: 3
• thesis: 1

DOI completeness: 920 of 920 sources have a DOI or a no_doi_assigned: true marker. Missing DOI metadata: 0.

Claim provenance (trace-every-claim audit v1)

Every numerical claim in body prose on ingredient and product pages — matches the regex <number> <unit> where unit is one of ppb, ppm, mg/kg, µg/kg, µg/g, mg/L, µg/L, % — must trace to a citation in the same H2 section (before or after the claim is acceptable; sections are treated as logical units of provenance). Valid citations: [[sources/...]], [[regulations/...]] (for regulatory limits), [[testing/...]] (for analytical-method floors).

• Total claims scanned: 3845
• Supported: 3845 (100%)
• Unsupported: 0

By page type:

• ingredient: 1997/1997 (100%)
• product: 1848/1848 (100%)

Cochrane target: >95% supported. Current baseline reports the gap to that target as a follow-up queue. The v1 audit does not yet verify that the cited source actually contains the claimed value (that requires PDF reading per claim) — only that a citation exists. v2 will add PDF-side verification.

Detail: claim-provenance-audit.csv (per-claim records with section context).

What this dashboard does not yet measure

The audits above are the mechanical, automatable dimensions of defensibility. The wiki’s Cochrane-level target also requires audits that this dashboard does not yet produce:

• PDF-side verification. Trace-every-claim audit v2 should follow each [[sources/...]] link to the cited source page, then to the raw PDF in raw/markdown/<FM_id>/ or raw/manual-fetch/..., and verify the claimed value appears in the source. Today’s v1 audit only verifies a citation exists; not that the citation is correct.
• Synthesis reproducibility audit. Pick 5 populated contamination_profile cells. Re-derive the values from contributing sources from scratch (without consulting the existing page). Compare. Target: 100% reconstructible. Not yet automated; requires deep reading of contributing sources.
• Inter-rater reliability. Re-ingest 20 source pages with a second extractor blind to the existing pages. Compare values + metadata. Compute κ. Target: κ > 0.7 extraction, > 0.8 inclusion. Not yet operationalized.
• External peer review. Three independent reviewers (toxicology academic, brand RA veteran, plaintiff’s-side class-action attorney) read the wiki and report defects. The commissioning brief is drafted at external-review-brief and is ready for Karen to send; reviewers not yet engaged.

These four audits are the Cochrane-level qualifying audits. They are documented here as gaps in the harness so the audit dashboard does not present a falsely complete picture.

Audit harness provenance

This dashboard is generated by tools/build-audit-dashboard.mjs on every prebuild. Inputs:

• data/evidence/taxonomy-coverage-audit.csv (output of tools/taxonomy/audit-coverage.mjs)
• data/evidence/product_source_routing_audit.csv, routing_unresolved.csv, routing_malformed.csv (output of tools/evidence/build-routing-audit.mjs)
• data/evidence/routing-discipline-summary.json (output of tools/audit-routing-discipline.mjs)
• Live re-run of tools/audit-ingredient-mandatory-sections.mjs and tools/audit-product-mandatory-sections.mjs
• Direct read of wiki/sources/*.md frontmatter for evidence tier + DOI accounting