Skip to content

REPRODUCE_20261009.md

Latest commit

 

History

History
27 lines (17 loc) · 4.68 KB

REPRODUCE_20261009.md

File metadata and controls

27 lines (17 loc) · 4.68 KB
Raw

Reproduction and evidence

  1. Use the exact source/plan perstudy; do not combine recipes. Keep historical checkpoints/datasets and all failed attempts. Python3.12.14/Torch2.6.0+cpu/NumPy2.2.6/SciPy1.15.1/psutil6.1.1/Pillow11.3 reproduce cloud CPU studies. GPU rendering used localTorch2.6.0+cu124/ModernGL5.12/glcontext3/NVIDIA581.29 RTX3090; runtime fingerprints andprecision flags are archived.
  2. RAMavailable≥8GiB before heavy neural loading/training, twoCPUthreads and at mosttwo scientific executors. Use the shared GPUFIFO locally; never interrupt unrelated workloads. Do not rerun closed fits. Replays and summaries do not count as new training realizations.
  3. scripts/summarize_sprint_results.py independently recounts complete saved cohorts. scripts/replay_VIS07_functional.py replays saved models through indexedfunctionalconv operations without fitting. It preserves ten extended-logit tolerance failures. scripts/diagnose_VIS07_depth.py tests the original andfunctional paths on the fixed128-row subset in a same-host run; a successful subset does not erase the failed broader attempt.
  4. neuropixel/rendering/nca_gl.py is the original scalarRGBA32Ffragment backend. nca_gl_vector.py is the separately verified vector variant. nca_gl_compute.py is the advanced OpenGL4.6/SSBO/shared-memory candidate; its engineering status must be read from the exact receipt, not assumed from its existence. Benchmarks check output parity before timing and retain all seven workloads and bothCUDA comparators.
  5. docs/animations/manifest.json links every animation to actual data by hash. The trained trajectory is an actual GPU-rendered capture of registered seed240/fixed TRAIN class0, explicitly labelled and slowed for inspection. Its source is results/research/VIS07_learned_gpu/actual_rendered_train0_seed240_frames.npz; the earlier CPU animation is preserved separately. Display colours do not change model states or prove semantic causality. GPU fidelity proofs and benchmark rows are separate evidence.

OpenGL 4.6 compute execution passed 21 Torch-free checks across all seven archived workloads. Its five-method timing study is complete: 245 registered rows, 77 output gates, 143 original rows unchanged and zero repeated timing keys. The original RAM-floor interruption and queue timeout remain archived history. Read RENDER10 and its admitted recovery before using the compute backend. Workload-dependent point comparisons across interrupted sessions do not establish universal acceleration or independent replication.

The publication manifest maps all retained raw archives (not the unavailable item-5 CPU archive) to their original Gitblobs, source commits, sizes andSHA256. GitHub main receives original immutable blobs plus documentation; rawbytes are not regenerated to disguise discrepancies. CodeMIT; UCI digitdataCCBY4.0, Alpaydin/Kaynak1998 DOI10.24432/C50P49. No originaltest264compositions were opened by new studies.

Independent external replication has not yet occurred. This package makes reproduction possible; it does not certify an outside laboratory result or a Nobel-level contribution.

Final engineering update — 9 October 2026

The recovered RENDER10 cohort is complete:245 registered rows,77 output gates,143 original rows preserved exactly and0 repeated timing keys. OpenGL4.6 compute measures2.084x forC16/grid128 relative to the bestCUDA comparator and loses forC48/grid128; interruption separates sessions and qualifies interpretation. Live physical viewer integration passed27 trajectory gates plus52 HTTP checks; CPU/Mesa passed the same separate internal endpoint. RENDER11 originally retained 103/105 energy blocks after the RAM floor stopped it; its saved recovery is admitted below; no wall-energy claim follows. See RENDER10_results.md, VIS07_viewer.md and RENDER11_results.md.

Admitted sustained-energy recovery — 12 October 2026

Do not rerun the completed RENDER11 measurements. The original contains 103 rows; the admitted saved recovery contains 105 unique registered keys and preserves all original rows. Use the kit environment in E: to verify or summarize saved receipts:

& E:/Proyecto-concursos/NeuroPixel-env-completion/Scripts/python.exe -m completion_kit.energy summarize --root . --original results/research/RENDER11_energy/original_20261009A/receipt.json --recovery results/research/RENDER11_energy/recovered_20261009B/receipt.json

See results/research/COMPLETION20261012/energy_recovery_admission.json for original and recovered SHA-256 identities, archived commit 87c60227, matching sources/inputs and exact original-row preservation. This is saved board-energy evidence, not a new physical run, independent replication or wall-energy measurement.