Percentes calibration (SPEC §10), instrument 86e96ceb642b5449ec15830fc816d38ede1bad95 target http://10.0.0.122:8000, gauge vllm:num_requests_waiting at http://10.0.0.122:8000/metrics started 2026-09-16T19:42:05Z, finished 2026-09-16T20:19:50Z ramp 1 rate_rps seed scheduled completed errored censored goodput queue gate verdict 2 42 259 259 0 0 1.0000 0.000 pass pass 4 43 488 488 0 0 1.0000 0.000 pass pass 8 44 884 164 0 720 0.0283 0.633 pass FAIL: goodput 0.0283 below 0.99 4.4 45 528 528 0 0 1.0000 0.000 pass pass 4.8 46 519 519 0 0 1.0000 0.000 pass pass 5.2 47 622 622 0 0 0.7074 0.008 pass FAIL: goodput 0.7074 below 0.99 lambda_max 4.8 rps ramp 2 rate_rps seed scheduled completed errored censored goodput queue gate verdict 2 1042 275 275 0 0 1.0000 0.000 pass pass 4 1043 499 499 0 0 1.0000 0.000 pass pass 8 1044 1008 148 0 860 0.0000 5.967 pass FAIL: goodput 0.0000 below 0.99; waiting-queue mean 5.967 above 1.0 4.4 1045 562 562 0 0 1.0000 0.000 pass pass 4.8 1046 556 556 0 0 1.0000 0.000 pass pass 5.2 1047 618 618 0 0 0.7832 0.000 pass FAIL: goodput 0.7832 below 0.99 lambda_max 4.8 rps ramps 1 and 2 agreed within 10% of the larger; lambda_max is the lower (§10) lambda_max 4.8 rps, lambda_r 3.12 rps (0.65 x lambda_max) load.rate_rps for the 2-replica experiment: 6.24 rps (2 x lambda_r) independent reference (§5) at 2 x lambda_r, 60 s warm-up, 300 s measured, judged by no gate rate_rps seed scheduled completed errored censored goodput censored_rate queue 6.24 2042 1900 519 0 1381 0.0216 0.7268 0.037