Final v0.3.3 validation evidence [M] · actual executed 7a0c93ff4d19b1892067711d884739d246432344; original CPU-WRF reference. Strict24 PASS; raw full D6 FAIL; raw integrity FAIL-new-class; frozen integrity annex PASS. All six breaches pass the frozen CPU-pair spread annex; integrity annex passes. Strict24 · Full D6 · Raw integrity · Annex

W3P0227: all 73 hours, spatial deviation bands

All native stored cells, no masks; finite-first [M]. Blue: GPU-versus-original-CPU RMSE; orange: signed mean bias. Dark/light bands: spatial |error| p25–p75/p5–p95, not confidence intervals. Grey line: original-CPU IC-pair empirical RMSE where available. Ratio row: deviation RMSE / pair RMSE where defined; reference lines1/2. The50% rule applies only to frozen D6 exceedance frames; zero denominators stay undefined. Dashed/off-scale arrow: unchanged D6 limit; hourly rain tau0 is undefined and its accumulated limit is informational.

Actual7a 0227IC-P vs original CPU R0227, IC difference included. Explicitbe0 nativeBYTE bridge; no execution relabel or best-of-N primary replacement; rawFAIL andfrozenannex remain distinct.; measured tree 7a0c93ff4d19b1892067711d884739d246432344. All stored cells, no masks, finite-first. Blue line: GPU-versus-original-CPU RMSE. Orange: signed mean bias. Dark/light blue: spatial |error| p25–p75/p5–p95. Grey line: CPU IC-pair empirical RMSE where available. Third row: GPU/CPU RMSE divided by pair RMSE, references 1 and 2 (the 50% predictability rule). Missing/zero pair denominators are undefined; no artificial floor. Pair spread is context, not a confidence interval or gate waiver. Dashed/off-scale arrow: frozen D6 RMSE limit; hourly rain uses the accumulated limit as an informational reference.

Actual7a 0227IC-P vs original CPU R0227, IC difference included. Explicitbe0 nativeBYTE bridge; no execution relabel or best-of-N primary replacement; rawFAIL andfrozenannex remain distinct.

Frozen empirical spread: original CPU-WRF vs CPU P0227 (theta +/-1 full-theta float32 ULP). References1/2 are guides; frozen gates remain separate. Pair spread is context, not fidelity truth, confidence interval or waiver. Original CPU-WRF supplies every blue deviation.

Summary n=1: cross-case sample SD is unavailable. Final summaries will show all cases and pointwise mean ± one sample SD per domain.

Plotted numerical data and source hashes