DulyRun Wave 0 / report 2
FA4 numerical control
PASS for classification of the known defective and fixed numerical signatures on the selected kernel fixture.
Result: all three defective-source arms produced the expected defective signature. Both fixed-source arms produced the expected fixed signature.
Protocol
- Workload: FA4 paged decode kernel at the selected Gemma global-layer shape, without model weights.
- Sequence: A1, A2, B1, A3, B2 on one NVIDIA H100 NVL.
- Defective source SHA-256:
2d81c74558f84f541206abdf4d30be586d0444360d2bc87f7fcacaa24e85d901. - Fixed source SHA-256:
0d1d50442601b874f88761cb8349d34a5578993c4dd7707135be88fd76694728. - Fixed source commit:
6d11b3c1b6ca7dac0b111bb0db63ca16eeb19a46. - The seed, fixture, warmups, replay counts, environment, and GPU were held constant.
| Arm | Source state | Observed classification | Control ratio |
|---|---|---|---|
| A1 | Defective | Expected defective signature | 1.0318 |
| A2 | Defective | Expected defective signature | 1.0283 |
| B1 | Fixed | Expected fixed signature | 0.0602 |
| A3 | Defective | Expected defective signature | 1.0307 |
| B2 | Fixed | Expected fixed signature | 0.0595 |
QA
All eight scoped checks passed: exact primary sources reopened, source states pinned, upstream issue reproducer retained, actual environment recorded, five raw runs retained, protocol held constant, both paired controls and categorical A/A checked, and claims limited to control classification.
Limits: speed is not calibrated because GPU clocks were not attributable to each short CUDA-event interval. The recorded timing values are used only to classify this known control. No model output correctness, end-to-end inference, release safety, energy, or cost-per-request claim is made.