docs: cross-model adversarial review trial (Antigravity/Gemini vs Claude)

[vpciii]

Summary

Sets up a manual, global process for using a second, different model — Google Antigravity (Gemini 3) — to adversarially review work Claude produced, for both code reviews and design decisions. It complements the same-model Claude PR reviewer (ADR 0021): a different model has different blind spots, so it catches a class of defect the author's own model structurally shares.

Per your call: trial-first (no ADR yet) and portable prompt + practice only (no Antigravity config wiring).

What's here

• experiments/adversarial-review/adversary-prompt.md — the skeptic persona you paste into Antigravity/Gemini. Refute-don't-bless framing, grounded in the project's own artifacts, with separate Code review and Design decision blocks and a BLOCK / PUSH BACK / NO STRONG OBJECTION verdict.
• experiments/adversarial-review/README.md — the trial guide: how to run it manually (Antigravity IDE or agy CLI), a trial-log template so graduate-or-kill rests on evidence (signal vs. same-model+CI / false-positive rate / cost), and the graduation trigger → ADR with the automation rollout path (run-gemini-cli / agy headless) as an explicit ADR-gated next step.

Deliberately a trial, not a ratified practice

• Lives in a new experiments/ dir, not methodology.md's numbered practices or the decision guide.
• No merge gate, no adoption obligation for any project.
• Graduates to an ADR only if the log shows it earns its keep across >1 repo; otherwise the directory is deleted and the negative result recorded.

Repo map kept honest: README layout row + CHANGELOG [Unreleased], both marked trial / no-adoption.

🤖 Generated with Claude Code

[claude]

Claude finished @vpciii's task in 1m 48s —— View job

---

Cross-model adversarial review trial — review

• Read governing artifacts (ADRs 0001/0004/0005/0019/0021, methodology.md, README/CHANGELOG)
• Review the 4 changed files against this project's own standards
• Post findings

---

This is a clean, well-disciplined docs PR. It reads like it was written against this repo's own rules rather than generic best practice, and it holds up. No blocking issues.

What's right (against this project's standards):

• Reversible-by-default (ADR 0004) is honored structurally. Trial-first, lives in a quarantined experiments/ dir, no merge gate, explicit graduation-trigger and a kill path ("delete the directory, record the negative result in an ADR"). A defended "no" as a first-class outcome also matches the spirit of ADR 0019.
• Artifact honesty (ADR 0005) is kept. Both the README layout row and CHANGELOG.md [Unreleased] ride along in the same PR and are explicitly marked trial / no-adoption — no drift between the repo map and reality.
• Terms align with the glossary/ADRs. expand-contract, blast radius/reversibility, appetite-vs-scope, and the ADR-trigger framing all match methodology.md and ADR 0004 — no domain-term drift.
• Process: Conventional Commit (docs:), Keep-a-Changelog Added under [Unreleased], and the ../../adr/0021-... link resolves correctly from experiments/adversarial-review/.

Two low-severity notes (neither blocks):

1. The portable prompt's Inputs: lines are project-shaped and won't literally resolve when dogfooded on this repo — adversary-prompt.md:60-61 and :80-81 name spec.md, docs/glossary.md, CONTRIBUTING.md, architecture.md. This methodology repo's own layout is adr/ (not docs/adr/), methodology.md (not spec.md), templates/project-CONTRIBUTING.md, and no top-level glossary. The shared framing already says "the project's own artifacts" generically, which mirrors ADR 0021's "reads whichever governing artifacts a repo has" — so this is just the Inputs: enumeration being downstream-project-flavored. Consider a one-line "use whichever of these your repo has" caveat so the prompt is honest when pointed at this repo.

2. README.md:47 links to ./log.md, which doesn't exist yet (create-on-first-use). There's no markdown link-checker in CI (.github/workflows/ only has the two review workflows), so this won't fail a check — but a relative link to a not-yet-existing file is the kind of thing ADR 0005 cares about. The inline "create on first use" wording makes the intent clear, so this is borderline; flagging only for completeness.

One question worth the author answering (not a blocker): introducing a new top-level experiments/ directory is itself a small structural decision (the territory ADRs 0001/0005 govern). The PR makes a deliberate, well-argued case for trial-first, no ADR yet, and the reversibility + explicit graduation→ADR path is exactly the right way to defer that — so I read this as correctly handled rather than process drift. Noting it only because it's the one judgment call the change rests on.
· experiment/adversarial-review