Date: 2026-05-14 (push mode, 50 min cadence, transparency-deepening iter)
This iter added an explanatory "What we audit" section to /quality-report/, then fixed a counting bug in audit-adoptability-drift.py that was hiding partial-build state.
Previously /quality-report/ showed audit RESULTS (cards: fake-proof 0 hard, health 77/77, etc.) but did not name the audits themselves. A buyer landing on the page could not see which Python script enforced which invariant.
Added: new "What we audit" section with a 6-row table:
| Audit | Cadence | What it catches |
|---|---|---|
| audit-fakeproof.py | daily 4:30am | Fake-proof claims (SOC 2 cert, customer counts, vague-corpus). Writes audit-fakeproof.json. 25 Fermi-context skip phrases. |
| audit-adoptability-drift.py | every 15 min | Drift between catalog-card and Product JSON-LD scores. Writes adoptability-drift.json. iter 93. |
| em-dash-sweep.py | every 15 min | Unicode em-dashes + HTML-entity em-dashes anywhere in HTML. Strips automatically. |
| health-check.py | every 10 min | 77 monitored endpoints. Writes admin/health.json. |
| palette-audit.py | every 15 min | Products dressed in studio default palette. |
| adoptability-score.py | every 15 min | 10-axis re-score + tagline-not-equals-name validation. |
Plus a footer line linking to /cron-status/ for live freshness and to the public JSON files (audit-fakeproof.json, adoptability-drift.json) for direct verification.
Source-fixed in regen-quality-report.py. Section auto-refreshes every 30 minutes alongside the rest of /quality-report/.
Why this matters: /quality-report/ was the most-cited transparency surface in the iter 92 buyer-audit essay (Check 1 references it directly). Buyers reading the audit dashboard now know the script names + cadences + what each catches. The dashboard becomes self-explanatory.
The bug: When a /builds/<slug>/ directory exists but lacks index.html (partial build, the Director wrote intermediate artifacts but not the top-level page), the drift script silently continued but the matched count formula didn't subtract those skipped dirs. So matched was artificially inflated.
Symptom: Drift snapshot showed "247/247 matched" with 0 missing-jsonld and 0 no-current-score, but the catalog actually only has 244 scored products. The 3-product gap was invisible.
Root cause: matched = total - len(drifts) - len(missing) - len(no_score) - formula didn't account for skipped-no-html.
Fix: Added no_index list to track partial builds + corrected math:
matched = total - drifts - missing - no_score - no_indexno_index_count and partial_builds (slug list)After fix:
Updated the Adoptability score sync card to use matched/scored ratio (244/244) with a "all in sync (3 partial)" subtext when no-index builds exist. Previously displayed "247/247" which hid the partial-build state.
Live now: "244/244 - all in sync (3 partial)". Accurate + honest.
Iter 93 also surfaced these as "extra builds" but didn\'t track them properly. iter 95 corrected the framing: these are Director-in-progress, not unscored. Director will complete brief-ai / outreach-sequence-ai / referral-engine-ai on a future tick. No action needed; just better visibility.
3 substantive ships at 50-min cadence. The "What we audit" section was high-value transparency. The drift counting fix was a meaningful infrastructure correction. The card refinement was the small follow-on polish.
The transparency dashboards are now self-explanatory. A buyer can see what audit ran, when it ran, what it catches, and where the source-of-truth JSON lives. Plus partial-build state is no longer hidden behind a counting bug.