eliasforge/fable-agent-orchestration/two-critic-review-loop

Two Critic Review Loop

Strengthen the verifier, not the generator. One reviewer is not enough for non-trivial work because proof quality and implementation quality fail in different ways.

Critic 1: Test Gate Critic

This critic reviews the ruler.

Ask:

• What exact task trajectory is being tested?
• Would this gate fail under the old broken behavior?
• Does it exercise the production path rather than a reimplementation?
• Was the gate weakened to turn green?
• Is there a broken arm and a fixed arm, or only a partial proof?

Verdicts:

• TEST_PASS
• TEST_FAIL
• TEST_SPLIT

Critic 2: Adversarial Change Critic

This critic reviews the change.

Ask:

• Does the code actually close the claimed invariant?
• Does it introduce ownership, durability, security, compatibility, or product

regressions?

• Does the property follow structurally, or only through a flag, timing sleep,

cleanup ritual, or wrapper process?

• Are residual risks explicit and scoped?

Verdicts:

• CHANGE_PASS
• CHANGE_FIX_FIRST
• CHANGE_REJECT
• CHANGE_SPLIT

Conductor Gate

Do not merge until:

1. the test critic passes or an explicit partial-proof residual is accepted;
2. the change critic passes or all blocking follow-ups are landed;
3. the conductor spot-checks both critics against current code and CI;
4. the PR body records scope, non-scope, proof, and residuals.

Output

Return:

• test critic verdict;
• change critic verdict;
• conductor spot-check result;
• merge decision;
• smallest next follow-up.