Skip to content
Dreadnode Dreadnode Dreadnode

Judges

Run a judge over an agent trajectory or other evidence.

Terminal window
$ dn judge <command>

Run a judge over an agent trajectory or other evidence.

outcome

Section titled “outcome”
Terminal window
$ dn judge outcome <--config> <path> <--output> <path>

Evaluate a serialized trajectory against a configured outcome judge.

Always writes an envelope JSON to --output. Exits 0 on success (envelope’s ok field is true). Exits non-zero on any error (envelope’s ok field is false). The envelope is the source of truth for the platform; the exit code is defense-in-depth.

Options

  • --config (Required) — Path to the judge config JSON (JudgeConfig).
  • --output (Required) — Path to write the wrapped judgement envelope JSON.
  • --trajectory — Path to the serialized trajectory JSON. Mutually exclusive with —session-id. Omit both for filesystem-only judges that don’t need a trajectory.
  • --session-id — Session UUID to fetch the trajectory for via the platform API. Mutually exclusive with —trajectory. Requires DREADNODE_SERVER, DREADNODE_API_KEY, DREADNODE_ORGANIZATION, and DREADNODE_WORKSPACE in the environment.
  • --include-compacted, --no-include-compacted (default True) — When fetching by —session-id, include pre-compaction history so the judge sees the full trajectory. Defaults to True for the judge use case.
  • --working-dir — The agent’s working directory inside the sandbox. Optional.
  • --input-format (default native) — Shape of the trajectory file (only applies to —trajectory mode). ‘native’ is Trajectory.to_dict() (notebook/CLI users); ‘openai’ is a flat OpenAI Chat Completions messages array (platform-side verification path). [choices: native, openai]