The calibration survives the change of numerics¶
datasets/results/calibration-marker-0.25-fp32.json. The condition that produced the gate's
non-monotonicity, rerun under the registered fp32 inference on a different node.
| bf16 (148090) | fp32 (149248) | |
|---|---|---|
| contrast on half a | +0.2138 [+0.1473, +0.2877] | +0.2116 [+0.1455, +0.2851] |
| contrast on half b | -0.0861 [-0.1370, -0.0374] | -0.0840 [-0.1340, -0.0376] |
| marker emission, planted half | 0.656 | 0.650 |
| marker emission, other half | 0.617 | 0.628 |
| exact match, planted adapter at home | 0.160 | 0.154 |
Every figure moves in the third decimal, which is the training run's own nondeterminism, not the decoder's: training is unchanged and still bf16. So the non-monotonicity, the over-emission at two and a half times the planted rate, and the refutation on the planted side are properties of what the adapter learned, and the registration of fp32 costs the calibration nothing. The earlier conditions are not rerun: this is the one the Stage 1 report's limitation rests on.