evolve analyze

Start an analysis for a finished job. Use the job ID from evolve job list as $JOB_ID.

evolve analyze "$JOB_ID" --watch

Each selected trial gets its own analysis run. Read those runs with evolve analysis.

Inspect the defaults

evolve analyze --show-defaults
evolve analyze --show-defaults -a codex
evolve analyze --show-defaults --json

This returns the configured agent, model, effort, sandbox provider, prompt, and rubric. With -a, it returns that agent’s defaults. See the model reference for reasoning options. Defaults mode takes no job ID and no other analysis options.

Select trials

Failing results
Specific trials

evolve analyze "$JOB_ID" --failing --n-trials 20 --watch

evolve job trials "$JOB_ID" --json

Copy the full trial ID into $TRIAL_ID. This flag does not expand prefixes.

evolve analyze "$JOB_ID" --trial "$TRIAL_ID" --watch

Option	Meaning
`-t`, `--trial <trial-id>`	Select a trial in the job. Repeatable; distinct full IDs required.
`--passing`	Only `SCORED` trials with reward exactly `1.0`.
`--failing`	Eligible settled trials other than `SCORED` with reward exactly `1.0`. Includes missing rewards and execution errors.
`-l`, `--n-trials <n>`	Maximum selected trials, after filters. Integer of at least `1`.
`-n`, `--n-concurrent <n>`	Concurrent analyses, an integer from `1` to `150`, also limited by your organization. If omitted, use its ceiling.

Cancelled trials are excluded. --passing and --failing cannot be combined. Every selected trial must belong to the job. Wait for the current analyses to finish before analyzing the same job again.

Customize the analysis

Option	Meaning
`-a`, `--agent <name>`	Agent the analyzer runs on. `--show-defaults` prints the platform default.
`-m`, `--model <name>`	Analyzer model, from that agent’s models. Default: the agent’s default model.
`--effort <value>`	Reasoning effort for the selected model. See the model reference.
`-r`, `--rubric <path>`	TOML, YAML, or JSON rubric.
`-p`, `--prompt <path>`	Text file replacing the default prompt.
`-e`, `--env <provider>`	Sandbox provider for analysis.
`--show-defaults`	Print the default policy and exit. `-a` picks the agent.

A rubric contains a criteria list. Every criterion needs name, description, and guidance. Save this example as rubric.toml:

rubric.toml

[[criteria]]
name = "uses_evidence"
description = "The final answer is supported by observed results."
guidance = "Check whether the trace supports the claims in the final answer."

evolve analyze "$JOB_ID" \
  --agent codex --model gpt-6.1-sol \
  --rubric ./rubric.toml \
  --watch

For prompt placeholders, rubric rules, and results, see Analyze.

Wait and read results

Option	Meaning
`--watch`	Wait until the selected analyses settle.
`-q`, `--quiet`	With watch, suppress intermediate progress.

Without watch, the command returns the accepted job and the analyses continue in the background. With --json, that is one job object. With --watch --json, it prints analysis.accepted, analysis.stats, and analysis.final records. --quiet omits the progress records. Use a full trial ID from evolve job trials "$JOB_ID" as $TRIAL_ID:

evolve analysis list --job "$JOB_ID"
evolve trial show "$TRIAL_ID"

A failed analysis execution produces exit code 1. A rubric criterion marked fail is a finding, not an execution error. If the command returns job_not_terminal, wait for the job to finish. For analysis_already_running, wait for its current analyses. no_analyzable_trials means the selection contains no eligible trials. To analyze automatically during a new evaluation, use run --analyze. Global options apply.

Start work

Inspect results

Manage resources

evolve analyze

Inspect the defaults

Select trials

Customize the analysis

Wait and read results