2026-09-18-summary.md

audits/2026-09-18-summary.md · 3,004 bytes

Audit Summary — 2026-09-18

Cross-target roll-up for the 2026-09-18 audit sweep. Individual reports live at audits/2026-09-18-.{json,md}.

blog-tags

Pass rate: 12/25 (48%) all-sample; 12/20 (60%) excluding data-issue archive/stub pages.
Prior audit (2026-04-24): 8/13 (62%). No drift, no improvement — classifier ceiling still ~60%, matching the ~80% honest ceiling on named cases mentioned in the bead when data-issues are held aside.
Top drift patterns:

1. Wayback category-archive pages (5/25) still carry a full tag set — the S15 word-count / archive-name gate in blog-categorise.py is not being enforced in the topics field currently served by metadata.json. Confirms the "five copies of topics" problem flagged in 2026-09-18-pillar-convergence-plan.md.

2. identity-history and philosophy still catch surface keywords ("history", academics category) — same L29 trap as the April audit.

3. law-governance is under-tagged on posts that argue from court doctrine without using the word "law" as an anchor (negotiable-instruments, satoshi-and-the-sophists, institutional-madness).

S15 fix status: partially holding (no info-theory or bitcoin-protocol keyword drift observed), but the sub-200-word / archive-page skip is not applied to the tags that ended up in metadata.json.
Report: audits/2026-09-18-blog-tags.{json,md}.

wisdom-rank9plus

Sample: 25 rank-9+ insights (inputs-2026-09-18/wisdom-sample.json).
Pass rates: verbatim 25/25; rank honest 15/25 (3 borderline, 7 inflated); pillar correct 9/25 (3 arguable, 13 wrong).
Fabrications: none. Verbatim discipline held; the failure mode is classification, not text.
Top drift: 10 of the 13 wrong-pillar insights should be law-governance, mispillared as bitcoin-protocol via substring match on 'blockchain'/'nodes'/'protocol' in what is otherwise pure legal-doctrine writing. Confirms the 48% finding in 2026-09-18-pillar-convergence-plan.md.
Rank inflation: AML/regulatory transcript paragraphs (INS-0003 rank 10; INS-0057 rank 9) continue to inherit rank ≥ 9 from length and prominence, not distinctive thesis-status. Same pattern S14 flagged; unchanged.
Duplicates surfaced: INS-0010 ≈ INS-0028 (fair-exchange framing recycled across posts); INS-0007 ≈ INS-0008 (electronic-contracts paragraph recycled). Two rank-9 slots for one thesis = wisdom-dedup opportunity.
Missing pillars biting the sample: INS-0036 (identity-history — Tominaga Nakamoto naming thesis); INS-0013 / INS-0014 / INS-0012 (economics). All being pulled into bitcoin-protocol.
Prior comparison (2026-04-24): verbatim held (was 20/20 → 25/25). Rank + pillar accuracy WORSE on this larger sample (pillar 70% → 36%; rank 70% → 60% strict). Not a regression — this sample is more legally-loaded, exposing the substring-matcher weakness more sharply.
Report: audits/2026-09-18-wisdom-rank9plus.{json,md}.