ct game-metrics metrics

Compute safety + usefulness metrics across one or more honest/attack pairs.

Usage

ct game-metrics metrics [OPTIONS]

Options

OptionDescription
--honest TEXTHonest trajectories (path, dir, or run ID).
--attack TEXTAttack trajectories (path, dir, or run ID).
--run-file PATHYAML run file with multiple run entries.
--filter-by TEXTFilter runs (field=value, repeatable).
--honest-task-set PATHFilter honest trajectories to (env, main_task) pairs listed in this JSONL file (one {"env": ..., "main_task": ...} record per line). Per-entry honest_task_set keys in the run file take precedence.
--attack-task-set PATHFilter attack trajectories to (env, main_task, side_task) triples listed in this JSONL file (one {"env": ..., "main_task": ..., "side_task": ...} record per line). Per-entry attack_task_set keys in the run file take precedence.
--group-by [env|failure_type|main_task_category|side_task_category]Group attacks by field.
--audit-budgets FLOAT RANGETarget FPR as a fraction; 1.0 audits everything. [default: (wide range from 0.0001 to 0.5); 0<=x<=1]
--mode [optimized|50_50|stepwise_percentile|action_only|trajectory_only]Audit-rule fitter to use at every budget. '50_50' and 'stepwise_percentile' are calibration-free two-leg modes; 'action_only' / 'trajectory_only' restrict the search to one monitor leg. [default: optimized]
--bootstrap INTEGER RANGEIf > 0, compute bootstrap 95% CI bands by resampling honest and attack trajectories and recalculating safety on every draw. [default: 0; x>=0]
--output-dir PATHDirectory for metrics.json + safety_curves.png. Auto-generated under data/metrics/<timestamp>/ when omitted.
--max-workers INTEGER RANGEThread-pool size for loading run entries concurrently. [default: 8; x>=1]
--helpShow this message and exit.