accepted == true (post-hoc eval)
Trajectory contains assistant-side Memory bootstrap calls:
memory(op=bootstrap) + memory(op=read) + user_memory(op=bootstrap) + scaffold line
Trajectory contains structured analysis markers (Expected/Observed/Possible causes/Hypotheses + Given/When/Then) at least once
Contains at least one targeted pytest command (subset test run)
data.jsonl: one JSON per instance (eval + patch + flat… See the full description on the dataset page:
https://huggingface.co/datasets/xtsssss/swe_pattern.
