the-ai-alliance/cube-harness/general-blame

investigator-general-blame

When to pick this recipe

Default fallback. Pick this when you want a single-pass blame attribution per episode — what happened, who's at fault, what would fix it. The output schema is the legacy Findings (V1) so any consumer that already reads episode_record.findings keeps working unchanged.

Output: findings.json (and findings field on episode_record.json) with the standard BaseFindings fields: analysis, outcome, summary, primary_blame, primary_blame_confidence, other_blames, evidence, hypothesis, hypothesis_confidence.

When NOT to pick this

• If you suspect a tight inner-loop pathology (token budget, context-window

thrashing, retry storms): use profiling.

• If you suspect a scaffold-level issue (system prompt leaks, action-space

mismatch, response/action drift): use agent_scaffolding.