Results

Each evaluation writes to outputs/eval/<experiment>/<trial>/:

summary.json — aggregated stats (pass ratio, reward mean/std, error count)
results.csv — one row per trajectory with the fields below
evaluated_tasks.csv — wide format task_id, traj_0, ..., traj_{N-1} consumed by seta_env.dataset.filter_tasks; see evaluation.md
trials/ — per-task agent logs and perf traces

Per-trajectory fields (results.csv)

Field	Meaning
`task_name`	Task identifier
`traj_i`	Trajectory index (0 to n_trajs-1)
`uid`	Unique session ID for this trajectory
`reward`	Reward from verifier (0.0–1.0, null on error)
`error`	Whether this trajectory hit an error
`error_stage`	Which stage failed: `1_reset_env`, `2_run_agent`, `3_evaluate`, `4_calculate_reward`
`error_message`	Error details
`iteration_count`	How many agent turns ran
`termination_reason`	Why the agent stopped: `task_finished`, `max_iteration_reached`, `max_tokens_exceeded`, `max_parse_errors`, `step_timeout`
`total_tool_calls`	Total tool calls made across all turns
`max_parallel_tool_call`	Largest batch of parallel tool calls in a single turn
`parse_error_count`	Times the model output couldn't be parsed as a tool call
`prompt_tokens`	Cumulative prompt tokens sent to the model
`completion_tokens`	Cumulative completion tokens generated
`total_tokens`	prompt_tokens + completion_tokens