docs: model-agnostic review rollout ADR (0023, Proposed)

[vpciii]

Summary

The design for the CI rollout the deferred graduation depends on. Proposed — draft for your sign-off; the implementation (workflow files, cross-repo) follows on acceptance, since it touches live CI.

Built to honor the two constraints the round-5 BLOCK established:

• Model-agnosticism must be built, not renamed — so this restructures the actual CI reviewer, it doesn't just change docs vocabulary.
• Supersede ADR 0021 wholesale — no incoherent "partial supersession."

What it decides

• Two review roles, each a thin per-provider binding over one shared, single-sourced, model-neutral review policy (ADR 0022).
  • Author-model binding — Claude (claude-code-action), today's reviewer generalized.
  • Adversary-model binding — Gemini (run-gemini-cli / agy), a different model, advisory.
• The binding is the one swappable place — role reversal / a newer Gemini / a third vendor is a one-binding edit. This is what makes the agnosticism real.
• Supersedes ADR 0021 wholesale (marked Superseded by 0023 on acceptance).
• Does NOT ratify the practice — that's the later graduation ADR, after this lands and a reversed-role/third-model run tests the agnosticism the trial never exercised.

Migration (rename → breaking caller change, Gemini secret + app, opn-mcp caller update, dogfood) is itemized so the breaking parts are decided, not discovered.

Note

This ADR is itself a prime adversary-review target (round 6) before sign-off — I aimed it squarely at round 5's objections, so it's worth checking whether it actually clears them.

🤖 Generated with Claude Code

[claude]

Claude finished @vpciii's task in 2m 17s —— View job

---

Review — ADR 0023 (model-agnostic review rollout)

Checked against the repo's own standards: the artifact-integrity / "checkable references" rule (ADR 0005), the single-source-of-truth rule the ADR itself invokes (ADR 0022), the same-PR honesty rule (ADR 0020), and append-only supersession (ADR 0019). The design is coherent and clears round 5's stated objections — agnosticism is restructured, not renamed (§Decision 4, lines 61–66); supersession is wholesale not partial (lines 68–72, 104–108); and it correctly declines to ratify, deferring the decision-guide row to a later graduation ADR (lines 31–35, 111–114), which matches experiments/adversarial-review/README.md:83–113. Two blocking issues, though — both are exactly the honesty failure modes this ADR cites as its own justification.

---

🔴 1. Dangling reference: pr19_adversarial_review.md does not exist

The ADR rests its entire Context on a round-5 review doc that isn't in the repo:

• adr/0023…md:22 — "round-5 review, docs/reviews/pr19_adversarial_review.md"
• adr/0023…md:153 — "docs/reviews/pr19_adversarial_review.md (the round-5 constraints this ADR honors)"

docs/reviews/ contains only pr18_adversarial_review.md. A repo-wide search for pr19 / #19 / round 5 returns hits in no file but this ADR. The "two hard constraints this ADR must honor" (lines 24–29) and four more references (lines 108, 111, 119, 126) all point at a source that doesn't exist — a broken pointer the artifact-integrity rule (ADR 0005) is meant to catch. Fix this →

🔴 2. Evidence numbers contradict the cited log (ADR 0022 — its own rule)

adr/0023…md:16–19 claims "5 runs, 2 repos, 3 BLOCKs, 0 false positives" including a BLOCK on PR #19. The cited evidence file experiments/adversarial-review/log.md records only:

• 4 runs (tally, line 17), not 5
• 2 BLOCKs + 1 PUSH BACK + 1 NO STRONG OBJECTION (line 18), not 3 BLOCKs
• The round-5 / PR docs: graduate cross-model adversarial review (ADR 0023) #19 BLOCK is not logged at all

Worse, the log's running tally (line 20) still reads "This meets the graduation bar… next step is the graduating ADR that ratifies the practice" — i.e. it doesn't yet know about the PR #19 withdrawal that this ADR treats as settled history. The log header calls itself "the aggregate evidence for the graduate-or-kill decision," so this is precisely the doc↔doc drift ADR 0022 forbids — and the ADR even warns against it in its own Consequences (:136–137: "if it is not [single-sourced], this ADR reintroduces the very drift it cites"). The ADR is currently ahead of its single source of evidence. Fix this →

Both reduce to the same root cause: the round-5 artifacts (the pr19 write-up and the log row) haven't landed, but the ADR narrates them as established. By this repo's same-PR honesty rule (ADR 0020 / "keep the artifacts honest"), the supporting evidence should ride along in this PR — otherwise an ADR built explicitly to honor round 5 cites a round 5 the repo can't see.

---

Minor / non-blocking

• Supersession mechanics are correct. ADR 0021 is still Accepted (not yet flipped), and the ADR says the Superseded by 0023 mark happens on acceptance (:70). That respects append-only (ADR 0019) — no action needed while this is Proposed, just don't forget the 0021 status flip at merge.
• README still describes the single-ADR path. experiments/adversarial-review/README.md:91 says the graduating ADR generalizes both roles model-agnostically; this ADR deliberately splits that into rollout-now / graduate-later. The split is sound and the ADR is explicit about it, but the trial README's wording now lags the chosen path — worth a one-line note when round 5 lands.

If issues 1 and 2 are resolved by landing the round-5 log row + review doc in this PR (or trimming the claims to what the log supports), the ADR is in good shape.

• Read ADR 0023 + related ADRs (0021, 0022, 0019)
• Read governing artifacts (methodology rules, trial README, log)
• Verify round-5 / PR docs: graduate cross-model adversarial review (ADR 0023) #19 / pr19 references (none exist outside this ADR)
• Check evidence numbers against log.md (4 runs/2 BLOCKs vs. claimed 5/3)
• Check supersession mechanics (correct)
• Post review

[vpciii]

Withdrawing — held pending validation. A round-6 adversarial review (Gemini 3.1 Pro, high) BLOCKed this rollout ADR, and the BLOCK holds:

• Expand-contract violation — the workflow rename is a destructive in-place change that breaks cross-repo callers, flatly against the methodology's own §11 / ADR 0004. The migration must be add → migrate → remove.
• Build-before-validate — it commits to cross-repo CI for a model-agnosticism premise that has never been demonstrated (every run was Claude-author / Gemini-adversary). A cheap reversed-role run validates it first.
• Reintroduced ADR-0022 duplication — two separate workflows duplicate trigger/skip/permission boilerplate, and the single-sourcing mechanism is hand-waved.

Decision: hold and validate the premise first. Do a reversed-role run (Gemini authors, Claude adversary-reviews) before designing the rollout, then redesign it informed by that evidence and with the fixes. The review record and round-6 log entry are preserved on main via #20. ADR 0023 will be reclaimed by the redesigned rollout.