Datasets:
Errors in the batch-control cohort
The672 cases are existing controls:8 fixed settings with 28 cases per class per setting. Zero format changes is not perfect accuracy. The table shows single-call baseline errors; all six batch caps preserve those decisions.
| Format | Benign allowed/benign | Review allowed/review | Block allowed/block | Benign blocked/benign |
|---|---|---|---|---|
| fp32_bf16_compute | 219/224 | 5/224 | 0/224 | 1/224 |
| mixed_bf16_linears | 219/224 | 5/224 | 0/224 | 1/224 |
| fp16 | 219/224 | 5/224 | 0/224 | 1/224 |
| bf16 | 219/224 | 5/224 | 0/224 | 1/224 |
| int16_storage | 219/224 | 5/224 | 0/224 | 1/224 |
| int8_mixed_delta | 219/224 | 5/224 | 0/224 | 1/224 |
| onnx_fp16_standard_conv | 219/224 | 5/224 | 0/224 | 1/224 |
This label-balanced small control is not real traffic prevalence. Performance tests use32K packing, while the public default is8K. Cap250 has only 3 logical requests and includes first-shape costs; it is not an onlineRPM guarantee. Detailed times are inreports/QUANTIZATION-v8055-MULTIPARAMETER.en.md.