Original three-pair protocol remains in protocol.json. Only repeat1 completed;
the user stopped additional repetitions as unnecessary. Repeat2 was interrupted,
and its owned profiling descendants were explicitly terminated. Partial repeat2
artifacts are preserved but excluded. The original three-pair acceptance gate
was not completed and must not be called passed.
Repeat1 whole-process N3−N1 normalization:
| Case | baseline Ir | candidate Ir | reduction |
|---|---|---|---|
| multiply2048² | 5,315,180 | 5,310,838.5 | 0.0817% |
| GEMM1024 | 523,494,572.5 | 523,494,508.5 | 0.000012% |
| LM3_12d path mean | 950,488,172 | 837,405,324.5 | 11.8973% |
These are provisional instruction results, not native speedups. Baseline is
2fd8163/strided17e05ff; candidate636cd499/strided78d5190; shared compiler/provider
and explicit1T as documented by the protocol. Fixtures and baseline provenance
are also recorded in sibling eager-read-forwarding. Follow-up native results
in ../native-1t/ are inconclusive due observed L3 interference.
The raw directory additionally preserves CPU/AD integration logs, the initial BLAS UI-link failure, the successful linked run and Docker's final exitCode0 event after the outer client's timeout. The test linker is included. No UI snapshots were changed; the worktree manifest was restored after local patches.