| Isolate protected regression |
analysis/protected-abba-2.json |
Complete matched A/B/B/A diagnostic population |
Does not replace prior 3% protected threshold or qualify performance |
| Check native state roundtrip |
analysis/state-logprob.json |
Reduced-geometry signed-package roundtrip passed |
No host offload, full logits, or exact-128 proof |
| Localize sampled distribution difference |
analysis/state-logprob.json |
Four top-20 observations retained |
Case 51 has MTP batch-shape confounder; no causal attribution |
| Observe scheduler behavior |
analysis/trace-paired-1/trace-paired-1-summary.json |
72/72 measured requests and 144 samples |
No HTTP-to-slot/generation mapping; graph mixtures differ; not request timing |
| Numerical qualification |
prior campaign and review/final-trace-causal-review.md |
Not covered |
Needs raw-logit or independent teacher-forced package |
| Promotion/default change |
summary.json |
Not covered and not authorized |
promoted=false |