CPU RESULT AUDIT: PASS

Independent standard-library CSV recomputation; no benchmark import, no inference.
Outer/timing/kit manifest entries: 31 / 18 / 65.
Four final checkpoint hashes; 42 runtime source hashes; 64 frozen official-train IDs.
Three complete processes, 64 pairs each, 384 measured system calls, all retained.
Host: AMD Ryzen 7 7840HS w/ Radeon 780M Graphics
Full-4: 764.292839583 +/- 4.982878236 ms (mean +/- process sample SD).
Stream-3: 575.700093229 +/- 4.506963933 ms.
Saving: 24.674256484% +/- 0.616790771 percentage points; range 23.962710784% to 25.056635577%.
Returned summary values match to absolute tolerance 1e-10 or relative tolerance 1e-12.
Tukey descriptive flags: Full-4 7; Stream-3 14; pair savings 13. No exclusions.
Pairs in which Stream-3 was slower or equal: 0 of 192.

LIMITS
- Recomputed from the returned CSVs; this audit is not a remote attestation of the host or an independent rerun.
- All 64 cohort IDs are in official train, and their reported original NPZ hashes match the pinned inventory. The NPZ bytes are not included in the return bundle; their user-side verification is evidenced by the pinned script and completion records.
- The protocol, cohort, script and runtime-source hashes match the supplied pre-run kit. Source review confirms CPU-only execution, eval/inference mode, no training, no label scoring, and no reused Full-4 logits.
- Three fresh processes on one host; the 192 pairs are repeated measurements of 64 inputs, not 192 independent datasets or hosts.
- Sample SD is between the three process means (or three process savings), not a confidence interval. No timing observations are removed.
- Resident-input, batch-one, one-thread CPU pipeline, including resize and modality generation but excluding disk read/NPZ decode, model load, process startup and pose estimation.
- CPU timing is a post-submission descriptive extension with final checkpoints; it does not remeasure or validate the historical RTX 4090 benchmark or alter the frozen accuracy gate.
- All four models remained loaded for both systems, so memory reduction was not measured.
- No CPU frequency, thermal/power trace, host-load trace or energy measurement is available. The reported 100 ns timer resolution is not accuracy or statistical precision.
- Report approximately 764.3 +/- 5.0 ms vs 575.7 +/- 4.5 ms and 24.67% +/- 0.62 percentage points; retain full precision only in machine-readable files.
