adversarial-defender

To defend a plan against rival AI models, use an adversarial defense loop that requires evidence-backed rebuttals and concrete plan diffs. You paste attacks from other models, and the defender generates structured responses citing specific evidence to refute them.

The TACT evidence bar enforces Truth, Authenticity, and Clarity requirements during adversarial review. It strictly forbids hallucinated evidence or invented citations, ensuring all rebuttals against model attacks are grounded in verifiable facts.

Structure multi-round rebuttals by maintaining an attack ledger that normalizes and classifies rival attacks as Valid, Partial, Invalid, or Needs-More-Info. Each response must steelman the attack, cite evidence, and emit an explicit plan_diff.

Yes, you must provide an exit condition along with the artifact, rival model, evidence standard, and environment access before submitting the first attack. These five Phase 1 inputs establish the boundaries for the structured defense loop.

The best way to steelman adversarial model attacks is to state the strongest version of the attack before rebutting it. This approach forces you to address the core argument directly and propose concrete, artifact-scoped changes via a plan diff.