Hugging Face's logo

PureOne
/
AUREOLE-R-v3

English
neural-rendering
computer-graphics
world-model
persistent-memory
ray-tracing
monte-carlo-rendering
control-variates
visibility-caching
active-sampling
uncertainty
reproducible-research
numpy
Model card Files
xet
 Community
AUREOLE-R-v3 / CHANGELOG.md
PureOne's picture
PureOne
AUREOLE-R 3.0.0-hf.1: standalone public research release
9d6c005 verified
preview code
|
Raw History Blame Contribute Delete
2.24 kB

Publication edition 3.0.0-hf.1

  • Added searchable Hub card and expert/AI-agent reading routes.
  • Added exact-result evidence selectors, citation formats, artifact inventory and synthetic-data dictionary.
  • Added hidden-token Windows launcher, account checks, byte-matching retries, parent-commit guard and anonymous public verification.
  • Preserved all scientific evidence and source; no new rendering experiment or performance claim.

v3.0.0 - Certified Innovation Rendering

Added six scoped results, geometric validity, residual-support elimination, sixteen tests and two new held-out protocols. Retained v2 mathematics, code, weights and adverse evidence. Updated standalone manuscript, API, agent metadata and publication launchers. No new model training, GPU execution or external publication.

Changes in 2.0.0

Version 1 supplied a broad world-model specification, ten scoped mathematical propositions, six synthetic witnesses and fourteen checks. Version 2 adds a working physical estimation loop rather than relabeling the original proposal as finished.

  • Add a frozen exact-integral correction API, with tests exposing same-sample refitting, wrong-integral bias and clipping bias.
  • Add canonical visibility memory, numeric checkpoint persistence, namespace checks, trust epochs and observation-triggered revision.
  • Train and package one small visibility prior, with explicit train/validation/test scene splits and a model card.
  • Execute a direct-light shadow benchmark on eight held-out scenes; retain active-sampling and stale-state failures.
  • Execute an independent eight-scene follow-up testing trust revocation without an oracle event flag.
  • Derive five additional scoped results: conditional correction, residual-risk information value, estimator equivalence, detection-delay limits and a fixed-budget concentration bound.
  • Package an installable NumPy API, physical demo, new tests, exact replay evidence, publication helper, full manuscript and original v1 archive.

There is no full SR/RR/FG implementation, GPU backend, game integration, broad calibration result or production speedup. Scientific maturity is estimated at 65%; the project remains classified as meaningful, not major or transformative.