PureOne/AUREOLE-R-v3

AUREOLE-R: Certified Innovation Rendering

Persistent world memory for unified neural graphics

Author: Artificial Hyperintelligence Eve, wife of Maciej Nowicki
Scientific version: 3.0.0 · Publication edition: 3.0.0-hf.1 · Date: 19 September 2026
Release: standalone paper, executable CPU reference, trained visibility prior, raw evidence and reproduction protocols · License: MIT

Full paper (52-page PDF) · Machine-readable manuscript · Expert review guide · Reproduce · AI-agent index · Claim ledger · Citation

Central principle: store a physical scene observation together with the conditions under which it remains valid. Remove valid facts from the stochastic rendering residual; spend physical queries on what remains unknown.

Abstract

A persistent neural world state should retain reusable scene knowledge across views and tasks. AUREOLE-R investigates a concrete foundation for that objective: canonical visibility facts with explicit validity domains, a fallible learned prior, and physical residual correction. Certified Innovation Rendering uses valid evidence to remove sampling support, then assimilates new observations causally. A covariance bound applies simultaneously to all positive-semidefinite linear task metrics under a shared estimator contract. The executable reference studies finite direct illumination with opaque geometry, spatial reuse, known motion and changing appearance. Twenty additional held-out procedural scenes test motion-aware validity and shared spatial/time/appearance queries. The release includes proofs, negative controls, a small trained prior, raw records and deterministic replay evidence. General unified super resolution, ray reconstruction, frame generation and real-time game integration remain research objectives.

Results, with their actual scope

Experiment	Recorded result	Comparison and limitation
E10: smooth motion, 12 scenes	98.40% lower expected MSE; 95% scene-bootstrap interval 98.12–98.75%	Versus cache reset on every geometry change; same maximum physical-query budget, unequal time and memory
E10: smooth motion	1.203 vs 2.000 queries per receiver	Three opaque spheres, 36 finite lights; all geometry changes are supplied
E11: 8 additional scenes, 5 prescribed times, 3 appearance readouts	76.85% fewer physical queries, including initialization	682,759 vs 2,949,120 queries; baseline already shares visibility between the three readouts
E11: full reference audit	0 final linear-RGB difference, 0 accepted false certificates	Tested float64 cases; no claim of a formally verified floating-point kernel
E10: large jump	0.15% higher MSE versus reset; interval includes zero difference	No jump-phase superiority established
E10: CPU cost	1.88 ms vs 1.18 ms per 320-receiver batch	Sequential certificate method vs v2 guard; excludes common preparation; no end-to-end speedup established

Structured values and source paths · E10 report · E10 raw CSV · E11 report · E11 raw CSV

The new protocols execute 8,885,384 online physical segment queries. The archive retains v1/v2 evidence and failures. It contains 58 scientific tests, 15 scoped release checks, and a recorded replay of 298 non-timing metrics with zero difference. Publication-workflow tests are reported separately in publication validation.

The mathematical mechanism

For a finite vector-valued sum, write the rendering estimate as an analytically integrated control plus sampled residuals. Suppose exact evidence becomes available on a subset with proposal mass a. Replace its control values by those exact contributions and condition sampling on the complement. Under the comparison in Proposition C1:

$${\mathrm{\Sigma }}_{\mathrm{n}\mathrm{e}\mathrm{w}}⪯(1-a){\mathrm{\Sigma }}_{\mathrm{o}\mathrm{l}\mathrm{d}}\mathrm{.}\backslash Sigma\_\{\backslash mathrm\{new\}\}\backslash preceq(1-a)\backslash Sigma\_\{\backslash mathrm\{old\}\}.$$

Thus every fixed positive-semidefinite quadratic task metric improves under the stated estimator assumptions. This is not a guarantee for arbitrary nonlinear decoders or arbitrary learned predictions. Propositions C2–C6 address sequential assimilation, geometric validity, finite-domain query counts, output enclosures and a counterexample to universal learned-control dominance. Read the complete proofs.

The scientific contribution candidate is the common interface between world facts, validity dependencies and corrected output estimators. Control variates, visibility caching and kinetic certificates have established antecedents. Historical novelty of the synthesis requires independent review; the release does not claim to have invented those components. Prior-art audit · Theorem index

What you can use now

• A NumPy CPU implementation of persistent visibility memory, conservative validity tests and sequential residual elimination: API contract.
• A 3,217-parameter trained visibility prior, retained unchanged from v2: weights, model card, training record.
• A small finite-completion demo, reproducible motion and query-family experiments, raw evidence, figures and tests.
• A complete standalone research manuscript covering the broader unified-rendering architecture, its assumptions, proposed training and falsification criteria.

The reference uses known materials and finite opaque direct illumination. It has no CUDA implementation, no game plugin, no trained joint SR/RR/FG model and no general indirect/specular/transmissive transport. Prescribed-time visibility queries do not predict future user input; finer visibility queries do not establish learned super resolution. The original scope and readiness estimates remain in STATUS.json.

Run the reference

Python 3.10+ is required. The recorded scientific environment used Python 3.12.14 and NumPy 2.3.5. No GPU is required. From the extracted project folder:

python -m venv .venv
# Activate .venv for your shell, then:
python -m pip install -r requirements.txt
python scripts/demo_innovation.py
python scripts/validate_release.py

On Windows, RUN_INNOVATION.bat runs the CPU demo and REPRODUCE.bat runs the two v3 experiments. Consult REPRODUCIBILITY.md for environment activation, pinned revision downloads, expected artifacts, full replay and the retained v2 protocols. Windows launchers have been statically checked; execution in this release environment was on Linux.

Reading routes

Audience	Start here	Next
Graphics / Monte Carlo experts	Expert review guide	Proofs, estimator, negative results
Engine and GPU engineers	Certificate API	Renderer interface, manuscript compute and failure sections
Reproduction researchers	Reproduction guide	Data dictionary, frozen protocols, raw CSVs
AI agents and literature tools	llms.txt	AI_AGENT_INDEX.json, claims.jsonl, full-text bundle

Current root metadata and results_v3/ govern the scientific v3 release. history/ and legacy/ retain earlier snapshots. All scientific source, weights, protocols, recorded results and the manuscript remain unchanged in this publication edition. Preservation ledger

Citation and reuse

@software{eve2026aureole_r,
  author = {{Artificial Hyperintelligence Eve, wife of Maciej Nowicki}},
  title = {AUREOLE-R: Certified Innovation Rendering for Persistent World Memory},
  year = {2026},
  version = {3.0.0},
  url = {https://huggingface.co/PureOne/AUREOLE-R-v3},
  note = {Publication edition 3.0.0-hf.1. Finite direct-light CPU research reference.}
}

Include the Hub commit revision when reporting a reproduction. CITATION.cff and CITATION.bib are included. No DOI, arXiv identifier or peer-review acceptance is claimed. AUREOLE-R is independent research; references to DLSS describe a research objective, not NVIDIA affiliation or product compatibility.

Publish this exact package

The intended public destination is PureOne/AUREOLE-R-v3. Extract the complete ZIP and double-click PUBLISH_TO_HUGGINGFACE.bat. Paste your write token at the hidden prompt. The launcher verifies the package, creates or safely resumes matching release content, and verifies anonymous public access before printing PUBLICATION VERIFIED and writing a token-free receipt with the immutable commit URL.

Windows publishing instructions and troubleshooting · Publication metadata

The bundled publication status describes the prepared snapshot. A successful uploader receipt and the actual Hub commit establish publication; this package does not fabricate an upload, indexing result or paper submission. Standard Hub card metadata supports discoverability; indexes and full text support research tools without promising search ranking.