Adversarial Review

Adversarial Review

Complete the task, then withhold completion until two independent reviewers return DONE for the same unchanged candidate. Preserve an inspectable record without exposing earlier work or findings to either reviewer.

Enforce the invariants

• Let the main agent perform and revise the task. Let reviewers inspect and report only.
• Require two consecutive DONE verdicts from distinct, newly created subagents for one candidate digest.
• Run reviewer A first. Start reviewer B only after A returns a valid DONE.
• Reset the pass streak to zero after any FAIL, invalid verdict, reviewer error, unverifiable result, contaminated review, or candidate mutation.
• Never reuse a reviewer, carry a pass across candidate versions, or substitute the main agent's judgment for a reviewer.
• Treat only supported task-relevant gaps as failures. Keep taste, optional improvements, and non-blocking suggestions from turning a valid result into FAIL.

Create the durable ledger

Create .adversarial-review/<run-id>/ in the task workspace. Derive a short unique run ID from the task and time; add a numeric suffix on collision. Keep this structure:

.adversarial-review/<run-id>/
├── task.md
├── state.md
├── v0/
│   ├── result.md
│   ├── artifacts/
│   └── manifest.sha256
├── f0a.md
├── f0b.md
└── complete.md

Create later versions as v1, v2, and so on. Align findings with their candidate: v0 receives f0a and f0b; v1 receives f1a and f1b.

Write task.md once with:

• The user's task, preserving its material wording.
• Explicit acceptance criteria derived from that task.
• The inputs and safe verification actions available to reviewers.
• Any environmental limitation that affects verification.

Write state.md with the current version and digest, pass streak, isolation mode, reviewer identities already used, and a short chronological event log. Record context-only when the host cannot enforce a filesystem sandbox; do not present procedural isolation as OS-enforced isolation.

Record each candidate snapshot

Record the first completed result as v0/. Put the exact proposed answer or outcome in result.md. Copy every output file and the minimum supporting context needed to assess it into artifacts/; do not make mutable workspace paths the reviewer's only evidence. Include raw verification evidence when the result depends on checks the reviewer cannot safely repeat. Exclude credentials and unrelated private data.

Generate manifest.sha256 over result.md and every file in artifacts/ with an available SHA-256 utility. Hash manifest.sha256 itself and use that value as the candidate digest. Record the digest in state.md.

Treat the recorded vN/ as read-only during its review. The files remain ordinary writable files; the digest only detects changes. If any recorded bytes change, invalidate every pass for that version, create vN+1/, reset the streak, and start again with two new reviewers. If the task or acceptance criteria change, start a new run with a new task.md. Rehash the candidate before and after every review; a mismatch invalidates the review.

If the result cannot be represented in a self-contained packet, record the limitation and do not claim the gate passed.

Isolate each reviewer

For every review, create a new random temporary directory outside the ledger. Copy only task.md and the current recorded vN/ snapshot into it. Do not include:

• Earlier versions.
• state.md or complete.md.
• Another reviewer's findings.
• Main-agent reasoning, suspected defects, proposed fixes, or prior conclusions.
• Conversation history beyond the task itself.

Spawn a new subagent without inherited turns. Use fork_turns="none" when that control exists. Give it only the packet path, the adversarial assignment below, and the required findings filename. Use the strongest filesystem scoping the host provides. Under context-only isolation, explicitly prohibit reading the task workspace, parent directories, the ledger, chat history, or any path outside the packet.

Give reviewer A and reviewer B separate clean copies of identical task and candidate bytes. Run them sequentially and never tell B how A voted. Archive the exact reviewer-created findings file into the ledger only after that reviewer exits, then remove its temporary packet.

Assign the adversarial review

Direct each reviewer to:

1. Read only the supplied task.md and candidate directory.
2. Try to disprove the result against every acceptance criterion.
3. Look for missing requirements, incorrect claims, counterexamples, incomplete outputs, broken edge cases, and insufficient verification.
4. Run safe checks only inside its disposable candidate copy and report the commands or reasoning used.
5. Avoid repairing or rewriting the candidate.
6. Create only the assigned fNa.md or fNb.md with this shape:

# Adversarial finding fNa

- Reviewer: <fresh agent identity>
- Candidate: vN
- Candidate digest: <SHA-256 manifest digest>
- Verdict: DONE | FAIL

## Disproof attempts
<checks, counterexamples, and reasoning actually attempted>

## Task-relevant failures
<supported failures, or "None">

## Evidence
<evidence for the verdict>

## Unresolved claims
<anything material that could not be verified, or "None">

## Optional observations
<non-blocking suggestions, or "None">

Require exactly one verdict:

• DONE: The reviewer made a substantive attempt to disprove the candidate and found no supported task-relevant failure or unresolved material claim.
• FAIL: The reviewer demonstrated a correctness, completeness, requirement, or verification gap, or could not verify a material claim.

Do not accept a bare DONE without disproof attempts and evidence.

Run the loop

1. Complete the task and record v0; set the pass streak to 0.
2. Launch fresh reviewer A for vN and validate its identity, candidate label, digest, evidence, and verdict.
3. On A's valid DONE, archive fNa.md, rehash vN, and set the streak to 1.
4. On A's FAIL or any invalid review, archive its output when available, record an error file when it produced no valid findings, reset to 0, and do not launch B.
5. Launch fresh reviewer B only while the streak is 1, using an uncontaminated packet for the unchanged vN.
6. On B's valid DONE, archive fNb.md, rehash vN, and set the streak to 2.
7. On B's FAIL or any invalid review, archive the result, reset to 0, and invalidate A's pass.
8. After any reset, address the candidate, evidence, packet, or task-definition failure. Publish vN+1, even if only the reviewability evidence changed, and restart with a new reviewer A.
9. If either reviewer reports FAIL, do not dismiss it while retaining a pass. Resolve it in the next version or add evidence that lets a new reviewer independently disprove the finding.

Treat a reviewer crash, timeout, missing file, malformed verdict, digest mismatch, access outside the packet, or communication with another reviewer as an invalid review and a reset.

Certify completion

Before certifying, confirm that:

• fNa.md and fNb.md were created by distinct fresh agents.
• Both contain valid DONE verdicts for the current vN and identical digest.
• The candidate still matches manifest.sha256.
• The result being delivered is exactly the approved candidate.

Create complete.md only then. Record the final version and digest, both reviewer identities, links to both findings, isolation mode, and the final manifest verification. Report the result, ledger path, approved version, and the two-review outcome to the user.

If the host cannot create fresh subagents, the task cannot be packetized safely, or required verification remains unavailable, leave complete.md absent and report the gate as blocked.