Audited results, negative findings, recovery tests, and explicit claim boundaries.
| Stage | Hard128 | MuSR64 | BBEH61 |
|---|---|---|---|
| Base | 99 | 39 | 33 |
| SFT | 103 | 44 | 36 |
| GRPO60 | 106 | 45 | 41 |
BBEH61 is single-seed; McNemar p=0.3018.
| Run | Tokens | Validation loss |
|---|---|---|
| Baseline | 16.384M | 4.755323 |
| Sweep-selected | 16.384M | 4.783076 |
The short-sweep selection did not beat baseline at the matched endpoint.
Exact: 22% → 36% → 47%
Format: 0% → 45% → 75%
LTX regression and bounded Hunyuan improvement are both disclosed. Human review remains pending.
Math · Pretraining · Video · Serving