docs: log adversarial round 8 (reverse defect-catch; bidirectional evidence)

[vpciii]

Round 8 is the data point rounds 5–7 were building toward: a reverse defect-catch.

Roles swapped and un-steered — Gemini authored the 6-practice expansion (methodology PR #22), Claude adversary-reviewed it cold, and this time the authored change had real defects: §15/§17 duplicate ADR 0014/0020 (violating ADR 0022), §13/§14 contradict the "no premature tooling" exclusion, §14 lacks deployable-service scoping, the set inverts scale-to-work, and it left the doc self-inconsistent. BLOCK.

Combined with rounds 1/4 (Gemini-adversary catching Claude-author defects), the trial now has independent defect-catches in both directions — genuine bidirectional support for model-agnosticism. Rounds 7+8 characterize the reverse direction fully: no false-positives on clean work (7), real catches on flawed work (8).

Residual: no third model yet; small same-week sample. The author-steering confound still applies to the design-ADR rounds (3/5/6).

🤖 Generated with Claude Code

[claude]

Claude finished @vpciii's task in 1m 55s —— View job

---

Review — adversarial round 8 log entry

• Read the PR diff
• Read repo standards (experiment README/trial guide, ADR index)
• Verify claims against the log's own framing + prior rows
• Check Conventional Commits / process drift
• Post review

Reviewed against this project's own artifacts — the trial guide (experiments/adversarial-review/README.md), the ADR index, and the log's internal conventions. This is a documentation-only change to the trial log; judged on accuracy and self-consistency, not generic style.

Verdict: clean, with one process-drift item to resolve before/with merge.

What checks out

• Tally arithmetic is correct. Runs 8 = opn-mcp ×4 (PR 13/14/15/16) + methodology ×4 (PR 18/19/21/22) ✓. Modes 3 code review (13/14/16) + 5 design decision (18/15/19/21/22) ✓. Verdicts 5 BLOCK + 1 PUSH BACK + 2 NO STRONG OBJECTION = 8 ✓. 1 retrospective + 7 pre-merge = 8 ✓.
• ADR references are real and correctly attributed — 0014 (refactoring guide), 0020 (user-facing docs ride-along), 0022 (single-source), 0018 (global summary). The §15/§17-duplicate-0014/0020 and ADR-0022-violation framing maps onto actual ADRs.
• The narrative is honestly hedged, which is the point of this log: the new tally bullet preserves the third-model gap, the small/same-week sample, the varied adversary tier, and keeps the author-steering confound attached to rounds 3/5/6. The "bidirectional support" claim is correctly scoped to the two independent directions (rounds 1/4 and round 8) rather than overclaiming from the steered rounds.
• Conventional Commits ✓ — docs: prefix, accurate scope.

One finding — dangling detail-doc reference (process drift)

experiments/adversarial-review/log.md:17 — the new row points to methodology docs/reviews/pr22_adversarial_review.md, but that file does not exist in this repo (on this branch or any branch). Every other methodology row resolves to a real file — pr18_, pr19_, pr21_adversarial_review.md are all present in docs/reviews/. Round 8 is the lone methodology row whose write-up is missing.

The README is explicit that "Detailed per-review write-ups live in the reviewed project (e.g. a docs/reviews/ file in that repo)" — and for a methodology PR, the reviewed project is this repo. So the log is asserting a primary-evidence artifact for its headline data point ("the reverse defect-catch") that isn't actually committed. For a log whose whole purpose is to be the auditable basis of a graduate-or-kill decision, the one row carrying the most weight shouldn't be the one with an unresolvable citation.

Recommend committing docs/reviews/pr22_adversarial_review.md (ideally in this PR, to keep the doc and the behavior it describes in the same change) before this lands. Fix this →

Minor (non-blocking)

• log.md:21 — the prose says "modes: 3 code review, 5 design decision" but the table header column is "Mode". Consistent enough; no change needed.

Nice work keeping the honest-confound framing intact while adding the new direction — the residuals section is the most valuable part of this entry.