Independent long energy replication of P8-S1
P8-S1 independently replicates energy noninferiority strongly enough to replace P8-S0 as the default.
Diagnostic work, exact algebra and local capabilities are not treated as end-to-end mining advantage.
What was tested?
P8-S1 independently replicates energy noninferiority strongly enough to replace P8-S0 as the default.
Why the test is meaningful
Promotion requires not only a pooled ratio but consistency across headers and a confidence interval entirely inside the preregistered engineering margin.
R_E=(J/hash)_S1/(J/hash)_S0gates: R_E≤1.001, headers≤1.002 on ≥6/8CI95_upper≤1.003How it was tested
Freeze S0/S1 and execute 64 long runs over eight new headers, one ABBA+BAAB pair each, with cumulative NVML energy, exactness and thermal audit.
What happened
S1 observation stayed 0.020378 s; kernel/process retention was 0.999687/0.999552. Energy/hash was 1.001967, only 4/8 headers met 1.002 and CI95 was 1.000160–1.003778. Maximum temperature was 66°C.
Exactness and statistical controls
The campaign was technically valid: exact B32/factors, zero discrepancies and at least 501 samples per process. All three registered energy-promotion gates failed.
What the result means
Full per-launch synchronization is preserved as a ~20 ms latency option, not the energy default. Repeating the same test is closed; a lighter primitive is required.
Limitations
- The negative verdict concerns default promotion, not correctness.
- Latency value may justify opt-in use.
- Results are specific to one GPU and driver/runtime stack.
Evidence trail
Use the sealed eight-header result to audit the three gates; new work must preregister a different synchronization mechanism.
Canonical variants
CANONICAL-EXP-158PREREGISTERED-CAMPAIGNSEALED-AUDITSource: internally audited canonical reports. Local filesystem structure, private headers and operational identifiers are excluded from publication.