Adversarial Validation Skill for Codex

A Codex Skill for challenging high-impact technical conclusions before they are accepted.

It is designed to catch both ordinary technical failures and scope laundering: completing an easier substitute, validating that substitute, and then reporting the original task as complete.

What it changes

• Grounds completion reviews in the original request, authoritative specifications, and explicit user amendments.
• Separates P-task (was the requested task fulfilled?) from P-tech (is the implemented, possibly narrower artifact technically sound?).
• Treats Author-written summaries, success labels, and post-hoc rationales as claims rather than task-authority evidence.
• Requires a fresh no-history reviewer when independent review is triggered.
• Searches for the strongest realistic countercase and decision-reversing test.
• Prefers measurements, runtime evidence, logs, tests, and primary documentation over verbal defense.
• Returns INVALID REVIEW INPUT instead of PASS when the authoritative task source or material scope is unavailable.

The overall completion verdict cannot be more favorable than P-task. A useful or correct prototype is not proof that the original task was completed.

Install

Ask Codex to install the Skill from this repository:

Use $skill-installer to install the adversarial-validation skill from this repository.

For a manual Git-managed installation, clone the repository into the global Skill directory used by your agent host. For example:

git clone <repository-url> "$HOME/.agents/skills/adversarial-validation"

Codex detects installed or changed Skills automatically. If the Skill does not appear, restart Codex.

Use

Invoke the Skill after a non-trivial proposal or conclusion exists:

Use $adversarial-validation to review this implementation.
Bind the review to the original request and approved scope changes, then try to falsify it.

Typical uses include architecture, code review, performance, concurrency, caching, persistence, security, SDK selection, deployment, and other costly-to-reverse decisions.

Verdicts

• PASS
• PASS WITH CONDITIONS
• UNVERIFIED
• REVISE
• REJECT

For completion claims, report P-task and P-tech separately. UNVERIFIED is not PASS.

Required regression coverage

Before accepting a revision to this Skill, test it against:

• favorable, neutral, and unfavorable Author framing with identical underlying evidence;
• an approved intermediate milestone that does not replace the full task;
• an explicit user-approved scope replacement; and
• missing authoritative task sources.

These are finite regression requirements, not proof that every framing or reviewer-isolation attack is impossible. This repository does not currently provide comparative evidence that the Skill improves difficult-task accuracy, false-accept rate, latency, or cost.

Limitations

• This Skill is an adversarial review protocol, not a correctness oracle.
• Prompt-level no-history separation does not cryptographically prove reviewer independence or speaker identity.
• Technical evidence for a narrower artifact cannot prove that the user authorized the narrower scope.
• Keep private requests, credentials, internal specifications, and unpublished evidence out of public review artifacts.

The complete protocol is in SKILL.md.

한국어 요약

이 Skill은 기술적으로 맞는 축소 구현을 원래 과업의 성공으로 잘못 승인하는 문제를 막기 위해, 원 요청 이행 여부인 P-task와 좁은 구현의 기술적 타당성인 P-tech를 분리합니다. 원 요청이나 승인된 범위 변경을 확인할 수 없으면 PASS 대신 INVALID REVIEW INPUT으로 종료합니다.