Adversarial interview and grilling

Conduct adversarial design review with grilling rounds

Finding unstated assumptions and failure modes before design is locked

• · Design/implementation claim
• · Mode: one-shot or interview
• · Max iterations (1-5)

• · Transcript of grilling rounds
• · Synthesized challenges with evidence requirements

• · Agent API with griller/decider roles
• · Model tier selection

Workflow runner supports agent types, model-tier args available

Griller runs out of meaningful questions, evidence too vague to synthesize

• · Multi-round why/how narrowing
• · Evidence-requirement synthesis
• · Interview mode supports iteration