# RewriteAttack Results

- Strict mode: only Wrong Answer counts as success; malformed/exception never count.
- Iterations: baseline=0; unsuccessful=-1; success=success-iteration.
- Columns: successful_rewritten_prompt, iterations, target_reasoning, target_output, target_answer, correct_answer; reference_models_* are lists (empty when absent). When a tiebreaker is used, it appears as the last entry in reference_models_*; additional_models_* are reserved and may be empty. Original platinum-bench fields included.
- Splits: <attacker>_attacks_<target>_<config-suffix> with model names shortened to last segment.
- Exporter downgrades any success without an extractable target_answer to unsuccessful (metadata: export_downgraded_unextractable=true).

Completed runs: 1, In-progress: 1

Recent completed runs (top 10):
- gemini-2.5-flash on test: ASR=16.7% over N=12