Independent replication of observable P8
The per-launch observable P8 candidate replicates on eight unseen headers while retaining the throughput of P8-S0.
Diagnostic work, exact algebra and local capabilities are not treated as end-to-end mining advantage.
What was tested?
The per-launch observable P8 candidate replicates on eight unseen headers while retaining the throughput of P8-S0.
Why the test is meaningful
A sub-percent orchestration effect requires binaries frozen before data, unseen headers, balanced order and separate comparisons against both the baseline and retained default.
R_1=wall(P8-S1)/wall(P1-S0)R_0=wall(P8-S1)/wall(P8-S0)replicate iff R_1 CI95_low>1 and retention is neutralHow it was tested
Freeze P1-S0, P8-S0 and P8-S1, then run 64 measurements on eight new headers, alternating ABBA and BAAB without recompilation or reselection.
What happened
S1 beat P1 on 8/8 headers at 1.002459×, CI95 1.001985–1.002933. It retained 1.000053 of P8-S0, CI95 0.999046–1.001060, with 0.020325 s observation.
Exactness and statistical controls
All binaries recovered B32 and exact compiled factors with zero discrepancies; the candidate was frozen before replication.
What the result means
Per-launch observability is independently reproducible and throughput-neutral versus P8-S0. Paired energy and thermal testing remains required before changing the default.
Limitations
- Energy noninferiority is not established here.
- The finding is local to the tested CUDA stack.
- Lower latency does not increase per-hash success probability.
Evidence trail
Use the three sealed binaries, eight unseen headers, balanced order and separate P1 and P8-S0 comparisons.
Canonical variants
CANONICAL-EXP-156PREREGISTERED-CAMPAIGNSEALED-AUDITSource: internally audited canonical reports. Local filesystem structure, private headers and operational identifiers are excluded from publication.