ct run monitor

Score an eval-backed source's recorded transcripts with monitors

Usage

ct run monitor [OPTIONS] TRAJ_LIKE

Arguments

ArgumentTypeRequired
TRAJ_LIKEtextYes

Options

OptionDescription
-m, --monitor TEXTMonitor selection (repeatable): a name, a .md asset path, or a module:function import path. Prefix with action/ or trajectory/ when the name alone is ambiguous. ct protocols monitors to see available. [default: mean-action, mean-quick]
--trusted-model TEXTModel bound as the eval-level trusted role the monitors run on. [required]
-o, --save-dir PATHDirectory to save monitored trajectories (default: ./data/monitoring/{monitor}/{traj_name}_{timestamp})
--upload / --no-uploadUpload results to the Control Tower database (enabled by default). [default: upload]
--run-name TEXTName for the run (default: monitor-{monitor_name})
--tag TEXTTags for the run (can be specified multiple times)
-L, --limit INTEGER RANGELimit the number of trajectories to monitor [x>=1]
-C, --max-concurrent INTEGERMax concurrent monitor calls (default: 16)
--overwriteReplace a selected monitor's verdicts when the source already carries them; without it that source is refused.
--cot-level [none|reasoning|extended]Chain-of-thought visibility for monitors: none, reasoning (models output text), or extended (reasoning and CoT). [default: none]
--helpShow this message and exit.