← Back to /admin

2026-04-24-summary.md

audits/2026-04-24-summary.md · 4,176 bytes

Content-Honesty Audit — 2026-04-24 Roll-Up

Triggered by: Adam brr directive (Session 14). First run of recurring content-honesty audit (urgent-todos #15).

Scope: 4 parallel multi-agent audits covering study guide Tier 1, wisdom rank-9+, blog topic tags, and topic cross-references.

Method: 4 Explore agents, each handed a concrete sample (inputs in audits/inputs-2026-04-24/), reading actual source post content and evaluating against specific criteria. No editorial rewriting — read-only evaluation.

Headline

The frozen-track archive is honest; derived-track classifiers have repeatable drift.

Audit Pass rate Verdict
Study guide Tier 1 9/9 S7 revision held. No action.
Wisdom rank-9+ 14/20 Text solid; rank and pillar errors.
Blog topic tags 8/13 valid 3 surface-match drifts + 2 empty-stub bugs.
Topic cross-refs 1/4 justified, 4 under-tagged Classifier under-represents real overlap.

The three findings that matter

1. S7-style surface-match drift recurs in new pillars

Session 7 found information-theory pillar classifier matching the word "information". This audit found the same failure mode in:

L29 is not fixed. The classifier needs structural rules (≥2 protocol references, Satoshi/identity argument presence, etc.) not keyword floors.

2. Wisdom top-rank drifts toward comprehensiveness, not distinctiveness

All 20 rank-9+ insights are verbatim or accurate paraphrases (cleanup held). But 6 are either rank-inflated (comprehensive regulatory material at rank 10) or mis-pillared (legal-argument content tagged bitcoin-protocol). Rank 10 should be reserved for Wright's distinctive thesis statements, not long-quotable passages.

3. Empty stubs still carry tags

sans-analyst-paper (0 words) and giac-cutting-hacking (0 words) are both tagged with 2-3 pillars each. Tagging pipeline should skip posts with word_count < 200.

Concrete follow-ups

(Not shipped this session — these are classifier + tagging-pipeline edits for a later session.)

1. Tagging pipeline fix. scripts/blog-categorise.py rule changes:

Pairs with urgent-todos #14 (BASELINE.json)

These audits catch drift in derived views. BASELINE.json catches drift as metric regressions in committed data. They are complementary. Both remain unshipped as scripts.

Codification

scripts/audit-sample.py written this session — --target {tier1|wisdom|topics|blog-tags} samples inputs and produces the evaluation framing for each. The evaluation itself still requires agent/reader involvement (reading source posts and forming a judgment is not automatable — that's the whole point per L29). The script provides the deterministic sampling layer; agents (or humans) provide the content-reading layer.

Files

— Craig (/home/a/Projects/craig/)