An open standard proposal — community-governed and MIT licensed, in the spirit of agents.md and Anthropic’s CLAUDE.md
ADVERSARIAL_REVIEW.md
AGENTS.md gave every repository a shared way to tell agents
how to work. ADVERSARIAL_REVIEW.md is the
open counterpart: a shared way to tell them how to doubt.
$ curl -O https://raw.githubusercontent.com/awdemos/adversarial-review/v1.0.1/ADVERSARIAL_REVIEW.md
# read it, edit the "Repo-specific rules" section, commitPinned to the v1.0.1 tag for provenance; verify with
git tag -v or the SHA-256 published in the
CHANGELOG.
Coding agent? Install the
adversarial-review skill
into your agent's skills dir and say run an adversarial review.
It is one file at the repo root that declares how adversarial review is done in that repository — which lenses, what counts as evidence, how severities are assigned — so any agent that reads it stops rubber-stamping its own output. Green CI proves the code does what its author imagined testing; this standard exists because agents now write code faster than humans can review it, and review has to be declared, evidence-backed, and machine-readable to keep up.
The standard is versioned, MIT licensed, and governed in the open. Changes go through a public process — issues before pull requests, cryptographically verified commits, recorded signoff on every merge — and anyone may participate. No vendor controls it, and no adopter must accept it unchanged: take the protocol, customize the repo-specific section, and propose improvements back to the community.
One file that makes review adversarial
- Drop an ADVERSARIAL_REVIEW.md in your repo root.
- It tells any coding agent: your work here is expected to survive adversarial review — and here is how that review is done in this repo.
- The agent reviews as an adversary, pass by pass, one lens at a time, with reproducible evidence for every finding and an explicit per-lens verdict. securitycorrectnessspec-complianceperformanceabuseoperationalgovernance
- Findings go back to the authoring agent; CRITICAL/HIGH items are fixed and the diff is re-reviewed before the work is called done.
- You customize the repo-specific section; the protocol and its six core lenses stay portable. (Changing the rules is itself a change — see the governance process.)
A prompt is a wish. A file is a rule.
-
Persistence
A prompt is forgotten; a file in the repo binds every future session — provided the agent is configured to read repo convention files, exactly the mechanism that makes
AGENTS.mdwork. -
Portability
Any agent that reads repo conventions picks it up with no integration; nothing about the standard is tied to one vendor’s tool.
-
Accountability
The protocol demands reproducible, severity-rated findings and explicit “checked and clean” verdicts per lens — the difference between a review and a vibe check.
-
Repo-owned rules
The repo declares its own invariants, known weak spots, and required checks. Generic review becomes this codebase’s review.
Canonical, versioned, copyable
The canonical, versioned template lives in this repo:
Review discipline, applied to itself
A standard about review discipline owes its own contributors the same
discipline: issues are opened and accepted before any PR, every commit is
cryptographically verified (SSH or GPG signed), and every merge to
main carries an NCO signoff — a named
non-commissioned officer of record who independently reviewed the change
and takes responsibility for it. Full process:
CONTRIBUTING.md.
Stable under semantic versioning v1.0.1
v1.0.1 — the protocol (lenses, severity taxonomy, evidence rule, output format) is stable under semantic versioning; template text may tighten in patch releases. v1.0.0 was battle-tested by a 347-agent adversarial review of this standard itself; its findings are in docs/review-2026-10-17.md. MIT License. Feedback and PRs are welcome under the governance process above.