ITSMBench

Trajectory viewer

4 models × 53 tasks × 4 trials. Every failed run shows the database diff the evaluator produced — the specific table, row and field where the agent’s end state parted from the expected one.

4 of the 6 models on the leaderboard have released trajectories; the rest are published figures only.