adversarial-review

Adversarial review prompts are generated by defining an artifact path, defender and rival models, attack angles, evidence standards, and exit conditions to produce paired defender and attacker prompts for multi-model evaluation.

Evidence symmetry enforces identical standards for Truth, Authenticity, and Clarity across both defender and attacker roles, ensuring both sides meet the same evidence bar to reduce rhetoric and unsourced claims.

A round-based protocol with plan diffs is enforced by iterating through attacker and defender rounds until a measurable exit condition is met, converging on concrete diffs and counter_proposal tags instead of vague commentary.

No, you cannot use different evidence standards. The protocol requires an identical TACT canon and forbidden-moves lists across roles, strictly enforcing evidence symmetry to prevent biased or unsourced claims.

To start an artifact stress-test, provide the artifact path, defender model and role, rival model(s) and attack angle(s), exact evidence standard, and the measurable exit condition for the loop.

No, a predefined measurable exit condition is required. The adversarial loop iterates by rounds, classifying claims and attacking or rationalizing them until the explicit exit condition is satisfied.