fix(agents): replace Critic scoring with claim audit

[FluffyAIcode]

Summary

• remove numeric Critic scoring and supportive-grader framing
• require full semantic thesis reconstruction and claim-by-claim adversarial audit
• require exact quoted claims, strongest objections, corrections, and residual uncertainty
• distinguish epistemic honesty from literal task completion
• explicitly audit numbers, awards, quantifiers, and claimed consequences
• tag benchmark stages with critic_protocol=adversarial_claim_audit_v1
• retain complete 923/923-token context with no sampling, summary, simplification, or fallback

Root cause

The prior Critic received the complete Generator response but was prompted to score it. Gemma therefore produced a stable supportive 10/10 judgment and overlooked material issues, including treating an unsolved problem as impossible, implying a Fields Medal is automatic, misstating the Clay prize in Chinese, and exaggerating downstream theorem consequences. Deterministic KV reuse correctly repeated that flawed protocol output.

Tests

• deterministic adversarial protocol requirements and full-context metadata
• targeted Agent GAN/report suite — 17 passed
• full platform-neutral suite — 867 passed, 1 skipped
• Python compile and diff checks pass

Made with Cursor

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.