Latency percentiles use positive latency samples from recorded model responses. Playback runs at a fixed 300 ms per frame.
Original recorded summary
{
"moves": 150,
"lines_or_food": 15,
"requests": 0,
"skipped_calls": 0,
"p50_latency_ms": 0.0,
"p95_latency_ms": 0.0,
"decisions_per_sec": 0,
"tokens": 0,
"cost": 0.0,
"mean_confidence": null,
"invalid_choices": 0,
"outcome": "step_limit",
"length": 17
}