Traces - Monte Platform

A trace is one run’s attempt at one task. An eval run writes one trace for each task it scores, so a score of 0.6 on 200 tasks rests on 200 traces. Traces show why a task passed or failed.

Trace Contents

Field	Description
`task_id`	The task that was attempted
`input`	The prompt: every message the model received
`output`	The model’s response, with each tool call and each later turn intact
`usage`	Tokens in and tokens out
`incomplete_details`	Why the response stopped early, for example the output token limit. Empty when it finished
`score`	The grader’s score, from 0 to 1
`expected_answer`	The gold answer from the task, when the task has one
`extracted_answer`	The answer the grader parsed from the response. Empty when the grader found none

A multi-turn or tool-using attempt is kept whole, in the message format of the environment. For example, a failed support-triage trace can show a search_kb call and an answer in output, with score 0, because the ticket had to go to a person through escalate_to_human.

View Traces in the Console

Click Evaluations in the sidebar.
Select the experiment, then the eval run.
Click a row in the task table to open its trace.

The table sorts by task order or by score. The sort and the page are part of the URL, so a link opens the same trace for another reader.

Download Traces

monte trace list support-triage/eval-03 --failed --sort score
monte trace download support-triage/eval-03 -o eval-03-traces.jsonl

Each line of the file is one trace with the fields above. For a training run with a grader, --step <n> keeps the rollouts of one step. Observability Commands lists the flags. To see what training changed on a task, open the same task ID in the baseline run, or use Compare Runs.