Native 1T exploratory comparison — INCONCLUSIVE

User authorized a distant idle L3 domain with workloads elsewhere. All four variants used CPU32, explicit kernel/provider1T, the same OpenBLAS0.3.26 shared library, three warmups and15 samples, sequentially. Thread checks and source inspection confirm RAYON_NUM_THREADS feeds CpuContext; OpenBLAS and PyTorch intra/inter-op getters report1. No thread-verification preload was used during timing (the GEMV experiment alone uses its diagnostic dispatch preload).

Both completed suites are INCONCLUSIVE under the predeclared interference gate: the first observed Julia and pi on the selected L3; the second observed pi. Julia affinity was0–63, so initial distance did not prevent migration. CPU24 and CPU40 preflights failed and no timing was collected there. A fresh all-domain preflight selected CPU32 again for the single complete rerun. No favorable case was selectively retried. The monitor also sometimes classified short-lived benchmark workers as foreign; the independently identified Julia/pi samples suffice to invalidate both suites regardless of those false positives. The executable-name heuristic subsequently added to run.py was not used in these recorded suites and does not retroactively change their classification.

LM opt_flops: raw medians (milliseconds), not accepted speedup evidence

Copy's apparent change reverses direction between suites (about3.9% slower vs 9.5% faster). GEMV's apparent incremental reduction is only about2.4–3.1%; its prior instruction result was4.3%, below the5% instruction gate. Neither trial justifies production promotion or an advertised speedup. GEMV remains a standalone diagnostic, with96 independent nonzero correctness checks passed.

All cases, both strategies, medians/IQRs, commands, exit statuses, timestamps, CPU activity and process observations are in summary.json, the raw logs and activity.json / runs.json; summarize.py regenerates the summary. Controls' identical second strategy reuses the first measurement and is not an independent sample. The existing harness retains aggregate median/IQR, not individual sample times. These are warmed eager-workflow times, including intrinsic dispatch and operand-handle work, not prepared standard-operation kernel-only latencies. Input creation, runtime/path setup, warmup and result destruction are outside its timer. Native inputs are zeros: separate unit/AD and nonzero probe validation must not be mislabeled a full nonzero LM cross-backend oracle.

No other workload was stopped or re-pinned, and no shared service was changed. The maintainer authorized temporarily excluding CPU32–39 from competing jobs, but the first affinity write failed with EPERM (the Julia jobs belonged to a different UID); passwordless sudo was unavailable. No measurement started. Independent verification found all 369 live-thread masks unchanged; see reserved/post-failure-verification.json. The attempted script's restoration warning concerned redundant writes to unchanged masks, not actual affinity loss. No privilege workaround or further timing retry was attempted.

The maintainer then stopped measurement attempts and requested PR integration. Previously measured instruction savings remain instruction results, not native speedup claims. Acceptance-quality timing and 4T remain deferred.