errors: []
incomplete: None
builds: {'A': '4e5b4ed1d5ac926f@Qwen3.8-27B-exl3-4.0bpw', 'B': 'd1d5ac926f@Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp'}

| prefill | conc | prefill tok/s A | prefill tok/s B | decode tok/s agg A | decode tok/s agg B | TPOT verdict | TTFT verdict | accept A/B |
|---|---|---|---|---|---|---|---|---|
| 1024 | 1 | 6930.2 [6874.2-7002.8] n=3 | 5869.6 [5840.1-5901.4] n=3 | 68.8 [68.4-69.4] n=3 | 75.7 [75.2-76.4] n=3 | WIN (-9.1%) | LOSS (+18.1%) | - / - |
| 16384 | 1 | 6174.4 [6101.1-6277.5] n=3 | 5483.8 [5442.7-5541.3] n=3 | 65.0 [64.6-65.3] n=3 | 71.1 [70.9-71.4] n=3 | WIN (-8.6%) | LOSS (+12.6%) | - / - |

fidelity applicable: True | each build was scored while it was the build the target was on -- the reference and its null control inside arm A's block, arm B inside its own -- over /v1/completions, on the same token sequence
null control: {"passed": true, "n_prompts": 2, "flips": 0, "max_kl": 0.0, "reason": "the baseline reproduces itself exactly, so a difference measured against it is attributable to the arm"}
row: {"prefill_tokens": 1024, "scored": true, "reason": null, "kl": 0.20046924058448864, "kl_positions": 12, "tvd": 0.1571031032975252, "missing_mass": 0.0022051407589587235, "argmax": {"flips": 19, "compared": 95, "flip_rate": 0.2}}
row: {"prefill_tokens": 16384, "scored": true, "reason": null, "kl": 0.18999174208270533, "kl_positions": 11, "tvd": 0.15117269667980812, "missing_mass": 0.004500030348728943, "argmax": {"flips": 17, "compared": 95, "flip_rate": 0.17894736842105263}}
pooled: {"prompts_scored": 2, "prompts_skipped": 0, "kl": {"mean": 0.1954582630401575, "lo": 0.18999174208270533, "hi": 0.20046924058448864, "clusters": 2, "n": 23}, "tvd": {"mean": 0.15413789998866664, "lo": 0.15117269667980812, "hi": 0.1571031032975252, "clusters": 2, "n": 190}, "flip": {"mean": 0.18947368421052632, "lo": 0.17894736842105263, "hi": 0.2, "clusters": 2, "n": 190}}
saved_as: measure_ab_1024-16384-build-exl3-mtp0-gguf-mtp0_20260923T210330Z_75b61d1a3ada3ecbf78486e16d0f490f8ac1aa20.json store_error: None
agreement A [(1024, 1, True), (16384, 1, True)]
agreement B [(1024, 1, True), (16384, 1, True)]
