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": 4,
"requests": 148,
"skipped_calls": 2,
"p50_latency_ms": 795.44,
"p95_latency_ms": 850.76,
"decisions_per_sec": 1.266,
"tokens": 102405,
"cost": 0.00430101,
"mean_confidence": 0.3657,
"invalid_choices": 0,
"outcome": "step_limit",
"length": 6
}