daniel/Artifex
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
Artifex/gpu_batch01_independent_review_v1_2.md
Daniel Maddern 203bc7917a Add GPU parity and feature gap analysis
2026-08-18 20:45:09 +07:00

1.2 KiB
Raw Blame History

GPU Batch01 Independent Review V1.2

Decision: REJECT

Exact Blockers

  1. All 97 features do not pass validation: calibration 94/97, holdout 46/97, and adversarial 67/97.
  2. Supertrend ATR limits were frozen as zero by the runner, not at the approved bounded values. The period-10 ATR calibration drift is 7.105427357601002e-15; the holdout drift is 1.4210854715202004e-14.
  3. Calibration-derived continuous limits do not generalize: holdout fails 48 Bollinger/Keltner requests plus three Supertrend ATR requests; adversarial fails all 30 Keltner requests.
  4. V1.2 contains no performance artifact or complete runtime provenance.
  5. ATR validation does not require ATR shape and NaN-mask equality before measuring finite errors.

Satisfied Checks

  • The contract contains 97 unique requests and was frozen before holdout/adversarial validation.
  • Reported finite numeric limits are mechanically enforced.
  • Donchian and PSAR outputs pass exact checks on supplied evidence.
  • Supertrend bands, predicates, direction, transitions, and final outputs pass exact checks.
  • Historical role semantics are honestly marked UNVERIFIABLE_FROM_RECOVERED_SOURCE and non-blocking.

GPU_FEATURE_PARITY_CONTRACT_V1_2_VALIDATED must not be issued.