compare-trajectories

To compare model runs, you need to align two trajectory exports sharing the same task set and harness. Trajectory diffing identifies per-task behavioral gaps and classifies divergence reasons like persistence, knowledge, or format parsing.

Trajectories diverge due to differences in strategy, retrieval, or decoding during model runs. Comparing aligned exports categorizes these behavioral gaps into persistence, recovery, knowledge, or format and parsing issues.

Trajectory diffing is the process of comparing two model run exports to pinpoint how and why their execution paths diverge on the same tasks. It produces an outcome-delta matrix and gap classifications for debugging.

Yes, trajectory comparison requires aligned run exports with a defined task-id space. Both model runs must share the same task set, harness, and tool-access mode to accurately identify per-task behavioral differences.

Yes, comparing trajectory exports highlights clean and actionable differences between model runs. These identified gaps are used to produce warm-start candidates for distillation or targeted model improvement.

Trajectory alignment only works when two model runs share the exact same task set, harness, and tool-access mode. It requires a trajectory-diff workflow and cannot compare runs with mismatched task-id spaces or differing tool access.