Mechanical Enforcement — How v7.6 Closed the Class B Gap

The 2026-04-21 Google Gemini 2.5 Pro independent audit triggered v7.5 (Data Integrity Framework, eight cooperating defenses, shipped 2026-04-24). v7.5 was a complete policy response — but most of its new defenses were Class B: they relied on the agent remembering to invoke them. v7.6 is the mechanical response. It promotes 4 silent agent-attention checks to write-time pre-commit failures, adds a per-PR review bot that fails the status check on new findings vs main, ships a weekly framework-status cron with a regression watcher, and explicitly enumerates the 5 gaps that cannot be promoted because pretending we could mechanize them would itself be a lie. This case study is the framework's full mechanical answer to the audit, published verbatim per the publish-then-remediate policy.

Read this first — outlier flag

This case study is itself an outlier in the corpus. v7.6 shipped in a single ~6-hour working session on 2026-04-25. Three biases stack:

Single-session execution — no organic cadence; phases ran sequentially in one sitting.
Dogfooded data collection — the author of the framework rework also wrote the data and reads it. Same-author confound.
Retroactive v6.0 application — the v6.0 measurement protocol shipped 2026-04-16; the data being reported is dominated by retroactive backfill, not by organic adoption on new feature work.

The full upstream case study labels the limits explicitly in §10 Outlier Limitations and applies them to the published numbers (e.g., 3.33 min/CU is a dogfooded micro-benchmark, not a generalizable velocity claim). Read the §10–§11 sections of the upstream before quoting any number.

Trust-page connection

This case study is the detailed mechanical answer to the Gemini audit, paired with the v7.5 policy answer. Together they are the framework's full reply to the 9 Tier 1/2/3 recommendations. The trust page Gemini-audit subroute links to both. Per the publish-verbatim policy: the original audit text remains unchanged on /trust; corrections and responses are appended.

Audit (verbatim): /trust/audits/2026-04-21-gemini
v7.5 policy response: the eight cooperating defenses (write-time + cycle-time + readout-time)
v7.6 mechanical response: the seven Class B → Class A promotions enumerated below + the five Class B gaps documented as unclosable-gaps.md

Summary card (T1 unless noted)

Framework version: v7.5 → v7.6
Trigger: Residual Class B → Class A gap left by v7.5; explicit user approval to "close the gap"
Ship sessions: 1 (2026-04-25)
Wall time: ~6 hours (T2 — Declared, single-session)
Phase 1 commit (0a23922): 4 new write-time check codes
Phase 2 commit (c0be8ea): PR review bot + history ledger + weekly cron
Phase 3 commit (ecb172d): Class B inventory + CLAUDE.md update
Phase 4 commit (58b82b5): manifest v7.6 bump + 616-line case study + propagation
New scripts: scripts/check-case-study-preflight.py
Extended scripts: scripts/check-state-schema.py (+2 check codes), scripts/measurement-adoption-report.py (+history)
New GitHub Actions workflows: .github/workflows/pr-integrity-check.yml, .github/workflows/framework-status-weekly.yml
Pipeline regression test: 8 → 15 assertions, all passing
Class A promotions in v7.6: 7
Class B gaps remaining (and individually justified): 5
v7.6 own state.json: instrumented end-to-end with v6.0 protocol from session-start (timing.session_start, cu_version=2, cache_hits[] populated, 6 contemporaneous log events)

The 7 Class B → Class A promotions

Concern	v7.5	v7.6
Phase transition w/ no log entry	Class B	Class A — `PHASE_TRANSITION_NO_LOG` pre-commit (1a)
Phase transition w/ no timing update	Class B	Class A — `PHASE_TRANSITION_NO_TIMING` pre-commit (1b)
Broken PR citation in case study	Class B	Class A — `BROKEN_PR_CITATION` write-time pre-commit (1c)
Case study missing tier tags	Class B	Class A — `CASE_STUDY_MISSING_TIER_TAGS` pre-commit (1d)
New findings vs main on a PR	Class B	Class A — `pm-framework/pr-integrity` per-PR status check (2a)
Append-only adoption history	Class B	Class A — dedup-by-date snapshot ledger (2b)
Measurement-adoption regression	Class B	Class A — weekly cron + regression issue (2c)

The 5 Class B gaps that cannot be promoted

Per docs/case-studies/meta-analysis/unclosable-gaps.md. Each gap has its own 4-section format (technical reason / observability / human action / tracking) in the upstream doc.

cache_hits[] writer-path adoption — the decision to recognize a cache hit is the judgment we cannot mechanize. Tracked at GitHub issue #140. Observable via make measurement-adoption.
cu_v2 factor correctness — magnitudes are judgment-based; we check presence, not whether novelty: 0.2 is the right number for this feature.
T1/T2/T3 tier tag correctness — preflight (Phase 1d) checks tag presence on post-2026-04-21 case studies. Whether the tag is the right tag (T1 vs T2 vs T3) requires reading prose in context.
Tier 2.1 real-provider auth checklist — Apple/Google sign-in handshake on a real device cannot be driven by an automated test runner without crossing into the mocking pattern v7.5 was built to avoid.
Tier 3.3 external replication — no pre-commit hook can simulate "an external operator on an unrelated product succeeded with the framework." This is the open invitation; see Gap 5 tracking.

Cooperating-defenses recap (v7.5 + v7.6)

 Write-time (pre-commit, fires in <5s):
   v7.5: SCHEMA_DRIFT, PR_NUMBER_UNRESOLVED
   v7.6: PHASE_TRANSITION_NO_LOG, PHASE_TRANSITION_NO_TIMING,
         BROKEN_PR_CITATION (write-time), CASE_STUDY_MISSING_TIER_TAGS

 Per-PR (fires on every push):
   v7.6: pm-framework/pr-integrity status check (delta vs origin/main)

 72h cycle (rear-guard safety net):
   v7.1 → v7.5: 12 check codes scanned across all features + case studies

 Weekly (trend signal):
   v7.6: framework-status cron (regression watcher on adoption history)

 On-demand readouts (any time):
   make documentation-debt | make measurement-adoption | make runtime-smoke

Tooling attribution (honest)

Per the publish-verbatim policy, the upstream §9 names every contributor with what each contributed. Summary:

Claude Opus 4.7 (1M context) — all v7.5 + v7.6 framework commits since 2026-04-21 carry the Co-Authored-By tag.
Google Gemini 2.5 Pro — independent audit on 2026-04-21 (different vendor, different model family, artifact-only access). The audit triggered v7.5 → v7.6.
OpenAI Codex — SSD audit on 2026-04-19 identified the dashboard build break and SSD sprawl that motivated several pre-v7.5 hardening commits. Per git log --since=2026-04-21 --pretty="%h %an %s", no commits in the v7.5/v7.6 window carry Codex attribution. The upstream tooling-attribution section explicitly leaves room to append further attribution if Codex work in this window is identified later.
Human (Regev) — trigger decisions, the four-part approval gate on 2026-04-25, policy choices (publish-verbatim, honest-status labels, Tier 3.3 sequencing).

What earned the v7.5 → v7.6 framework bump

A new structural capability — mechanical enforcement is a layer that did not exist in v7.5. v7.5 had write-time gates for schema and PR-resolution; v7.6 adds the transition checks (1a/1b) and the case-study checks (1c/1d), plus the per-PR + weekly recurring layer. These are not extensions of existing checks; they are new check classes.
Propagation across surfaces — manifest, CLAUDE.md, evolution doc, integrity README, repo-root mirrors, this MDX case study, and the trust page response section.
A measurement that the change is real — pipeline regression test extended from 8 to 15 assertions, all passing at every Phase 1/2/3 commit. v7.6's own state.json is instrumented end-to-end with v6.0 protocol — proof of concept that the protocol can be applied without retroactive backfill when started at session-start.

Lessons (excerpts — see upstream §14 for the full set)

Approval gates are multi-part. The user said "close the gap"; I executed Phase 1 immediately. I should have paused and explicitly answered all four sub-questions (class behavior, scope, version bump, Tier 3.3 sequencing). A new feedback memory captures this so it doesn't recur.
Write-time enforcement is cheaper than cycle-time enforcement when the cost is failure-mode latency. The 72h cycle is fast in absolute terms but slow relative to the rate at which a single agent can ship 5 PRs in an afternoon. Pre-commit fails in 3–5 seconds; the cycle catches the same class 0–72 hours later.
Class B is not a bug — but undocumented Class B is. v7.5 had 5+ silent Class B gaps that only surfaced when explicitly enumerated for v7.6. The act of enumerating them in unclosable-gaps.md is itself a v7.6 deliverable. A framework that knows what it cannot mechanize is more trustworthy than one that pretends every check is a check.

Mechanical Enforcement — How v7.6 Closed the Class B Gap from Gemini's Audit