WHOLE-LIBRARY CONTRADICTION & COHERENCE AUDIT

Regenerative Toronto | July 2, 2026

Document ID: MAIN-AUDIT-001 Scope: Cross-document coherence — same-fact-different-figure collisions, conclusions unsupported by the library's own cited evidence, confidence-framing inconsistencies, and claims that quietly conflict when read together. This audit checked documents against each other, complementing the per-document citation verification already tracked in PROCESS-002 and Tier 6. Method: (1) verification of leads surfaced during the full REF_001_Index read; (2) targeted wrong-variant sweeps for every locked Tier-1 figure; (3) automated keyword-clustered dollar-figure collision detection across all research documents; (4) direct resolution of each cluster the detector flagged. Baseline and final state: all automated tests passing (59/59 at start; 60/60 at close — one new regression test added by this audit).

FINDINGS — FIXED DIRECTLY (unambiguous errors)

F1. MAIN_REF_001_Index.md internal stats cluster (4 errors in one file)

The REF_001_Index carried three separate stats blocks in three different formats, only two of which the maintenance script knows how to refresh — the structural root cause of this file's recurring drift, not a one-off: - Line 17 said 72 documents / ~1.8MB directly beneath a correct, freshly-updated 613-claim line. → Fixed to 80 / ~2.7MB. - The tail status block said 28/28 tests passing (actual: 59/59, now 60/60). → Fixed. - The footer said Version 1.0 | June 30 despite multiple substantive July updates. → Bumped to 1.1 | July 2. Note: the footer sits at line ~258 with 50+ lines of content appended after it — the file has been extended past its own ending. Cosmetic, but recorded. - The "Next step" line still described the pre-audit plan ("external frontier-model review → … → Opus pass"). → Updated to the actual current sequence.

F2. The Ontario Works percentage coherence cluster — the audit's central finding

One underlying fact — OW as a share of minimum-wage income — exists in the library in three competing forms, and the library's own quality infrastructure was enforcing the wrong one: - BRIEF-001 (✅ sourced, CL-613, Open Policy Ontario): 26% (2025). - Five DERIV campaign assets + HMK-001:1329 + REF_002_Sources: ~24% (an internal calc on raw 40-hour assumptions; two DERIVs attribute it to "Feed Ontario 2019," which cannot source a 2025/26 figure). - HMK-001:1333 — the library's own explicit editorial rule, which everything else ignores: say "about a quarter"; do not use either precise figure. - Test 2.2 enforced the inverse of the verified ledger entry — it flagged any unhedged "26%" as "superseded by ~24%," i.e., the automated quality gate was defending the internal calculation against the sourced figure. It had only ever passed because BRIEF-001's phrasing happened to fall outside the regex's character window. - REF_002_Sources.md:742 contradicted itself in a single sentence: "…at $733/month against a $17.20/hour minimum wage, OW represents approximately ~24% … (on current $17.60/hr)" — two different minimum wages in one breath, dated "By 2026" when the verified figure is 2025.

Fixed: the REF_001_Index row (now 26% (2025), sourced, ✅); the REF_002_Sources self-contradicting sentence (now internally consistent, presenting both measures with correct attribution and the "about a quarter" editorial rule); test 2.2 (rewritten as a misattribution guard: fires on any Open-Policy-Ontario attribution of the internal calc figure, exempting correctly-labeled internal calc mentions). Flagged, not fixed (judgment call): which form the five DERIV campaign assets should carry — see "Flagged for human review" below.

F3. The $359 warming-centre figure — a documented refinement that never propagated

HMK-038 established (with explicit propagation guidance) that $359/night is the press-converged figure, while the AG's own chart reports a $336–491 actual range ($266–323 budgeted) — and warned that presenting $359 as the AG's literal number is "an easy, avoidable opening" for a hostile reader checking the primary chart. That refinement lived only in HMK-038 while HMK-001 (two locations), REF_002_Sources.md (two locations), and HMK-009 all still presented "$359 ✅ AG audit" unhedged — the precise vulnerability HMK-038 named, live in the library's most foundational document. Fixed in all five locations (presenting $359 as press-cited, consistent with the AG's own range), plus new regression test 2.22 locking the refinement per the standing manual-fix-needs-a-test instruction. Flagged, not fixed: DERIV-006's song lyric uses the bare $359 — creative copy is an editorial decision (below).

F4. TRACK_001_Master_Issues contradicting itself on the recommendation count

The Tier 2 intro carried a stale count (45, predating Gap #47's M-13/M-14 additions) while the same document's own header and Gap #47 both confirm the current count of 48 when M-13/M-14 were added. → Fixed, with an inline note.

F5. Model-name leakage — introduced and caught within this audit

My own first fix to the REF_001_Index "Next step" line included an AI model name in a public-facing document; test 7.1 caught it on the next run. Removed. Recorded here rather than silently, because it demonstrates the leakage test working exactly as designed — including against the auditor.

FINDINGS — CHECKED AND CLEAN (verified non-collisions, recorded so future audits don't re-chase them)

• TCHC $2.229B / $4.8B wrong variants: every remaining instance is correction-context (HMK-002, HMK-004, PROCESS-002 describing the fix). HMK-012's "$4.8B" is a different metric entirely (10-yr municipal NPV), correctly labeled.
• $13.8M: all instances are correction-context or a genuinely different fact (Calgary Homeless Foundation's $13.8M year-one allocation, HMK-011) — the "true number, wrong scope" pattern correctly kept separate. One micro-item: HMK-001:923's FOI verification-path list still references "$13.8M ⚠️" as the target figure; harmless (it describes what the FOI would check) but could name the corrected $13.2M — left as-is, noted here.
• $135 vs $136/night: not a collision — REF_002_Sources:99 explicitly reconciles them ($135 = winter-respite budgeted floor; $136 = base shelter benchmark).
• $1,058 vs $2,559 per admission: two different studies, two different populations (psychiatric-specific, Gaetz 2012 vs. all admissions, Hwang 2011), each correctly scoped inline where it appears.
• VSL ($6.5M 2007 / $9.8M 2026), FCI (13.4%→14.4%), Latimer ($6,311 / $58,972), IHAP ($261.87M / $97M), homeless counts (15,418 / ~12,180 / PIT 9,026): consistent everywhere checked; no stale variants found outside correction-context.
• OW "70%" vs "roughly 70-75%" (1990 baseline): not a contradiction — BRIEF-001's 70% is OPO-sourced (the index publishes the ratio directly); the "70-75%" hedge (standing instruction) applies to internal min-wage comparisons where the precise 1990 wage is untooled. Both are correct in their contexts. Noted so the difference isn't mistaken for drift.

KNOWN-TRACKED ITEMS DELIBERATELY NOT RE-FLAGGED

Already in TRACK_001_Master_Issues with owners/paths: #25 (HMK-001 §28C vs DERIV-007 table divergence), #50 (Dixon Hall $56K gap), #56 (TSSS 2025 $912.0M vs $897.957M — external-vs-primary, correctly hedged in place), Gaps #59–61, and the encampment four-metric finding (#46, resolved as non-contradictory). Re-listing tracked items as "findings" would inflate this audit; they are acknowledged, not duplicated.

FLAGGED FOR HUMAN REVIEW (judgment calls — not resolved by this audit)

1. OW figure canonicalization across the five DERIV campaign assets — RESOLVED, July 5, 2026. Confirmed still fully live during a Phase 2 pass on this exact audit document (three sessions after this item was first flagged, all five documents had already had individual citation passes that didn't catch this specific issue). Implemented the recommendation below across all five: DERIV-001, DERIV-002, DERIV-003, DERIV-005, and DERIV-013 now use "about a quarter" in campaign-facing prose with 26% (2025, Open Policy Ontario, CL-613) as the sourced reference figure, retiring the unhedged "~24%" internal calculation everywhere it appeared. Original recommendation, for reference: adopt HMK-001's rule — "about a quarter" in all campaign-facing copy, 26% (2025, OPO) as the sourced reference figure in research documents, retire the hours-assumption-sensitive ~24% except as a labeled internal calc.
2. DERIV-006's "$359" lyric. The album brief's theme line and a lyric cite the bare press figure. Creative copy tolerates imprecision differently than research documents — but the hostile-reader risk HMK-038 named applies to anything public. Your call whether the lyric stands (defensible as "consistent with the AG's range") or shifts to a range-safe phrasing.
3. "Poverty industry" framing — standing item, reaffirmed with precise locations. Live in HMK-001 (§5 header + TOC, load-bearing), HMK-015, HMK-038; deliberately avoided in HMK-037 — the inconsistency between HMK-001 and HMK-037 remains exactly as TRACK_001_Master_Issues describes. No changes made, per your standing instruction.

WHAT THIS AUDIT DID NOT COVER (honest scope limits)

• A claim-by-claim confidence-label sweep (every ESTABLISHED/REPORTED/INFERRED/SPECULATIVE tag re-justified against its evidence) — the framing checks here were driven by figure collisions and known pending items, not an exhaustive label audit.
• Re-derivation of every calculated figure (ROI chains, NPV, S-10's $3.0B derivation) — arithmetic spot-checks only where a collision suggested one.
• The keyword-clustered collision detector covers major recurring topics; a figure appearing in only one document, or under a keyword not in the cluster set, would not have been caught by that pass (though the locked-figure sweeps are independent of it).
• Per the project's own standing instruction: a spot-check is not exhaustive, even a careful one. This audit found real errors that previous "clean" passes missed; a future pass should assume the same about this one.

MAIN-AUDIT-001 | Version 1.0 | July 2, 2026 | Regenerative Toronto Fixed: 11 instances across 7 files, +1 new regression test. Flagged: 3 judgment calls. Verified clean: 6 figure families.