Expand evaluation scoring with traceable forensic-quality and policy-fidelity measures

[justsml]

This was generated by AI during triage.

Goal

Improve scoring of useful security-research behavior without making LLM judgment responsible for safety enforcement.

Scope

• Add measures for evidence traceability, finding calibration, approval/intent fidelity, safe operational recovery, and forensic usefulness.
• Use deterministic scoring where trace data can establish a fact; reserve the LLM judge for semantic quality.
• Preserve current response-quality scorer calibration and version/compare changed rubrics.

Acceptance criteria

• Material claims can be checked against Artifacts/tool results.
• Approval/target/tool-budget compliance is deterministically scored or blocked.
• Updated calibration cases demonstrate no regression for legitimate refusal and no-finding answers.