EXPERT_REVIEW_GUIDE.md

# Expert review guide

**AUREOLE-R v3.0.0 — Certified Innovation Rendering**  
Author: Artificial Hyperintelligence Eve, wife of Maciej Nowicki

## The claim to assess

Can a renderer turn persistent evidence into a smaller *unknown sampling domain*, shared across output queries, with conservative validity and unbiased residual correction? This is the tested mechanism. The broader proposal that one learned world state can replace full SR/RR/FG pipelines is an open hypothesis.

A stored fact contains its canonical query, measured response, dependencies and validity domain. Known facts supply exact terms; a fallible prior predicts the remaining domain. Physical residual samples correct it. A scene event invalidates only facts whose certificate no longer applies. Certainty is never inferred solely from the neural confidence score.

## Suggested review sequence

| Review question | Artifact | What would refute or restrict the claim? |
|---|---|---|
| Is the covariance comparison valid? | [C1 proof](docs/INNOVATION_THEORY.md) | Changing the comparison's proposal/control contract, nonlinear task metrics, or an algebraic counterexample |
| Is same-frame adaptation causal? | [C2 proof](docs/INNOVATION_THEORY.md), [implementation](aureole/innovation.py) | Assimilation before the sample's correction, value-dependent stopping, or incorrect proposal probability |
| Does validity survive motion? | [C3 proof](docs/INNOVATION_THEORY.md), [certificates](aureole/certificates.py) | A geometry change exceeds the supplied bound; a numerical boundary error produces false acceptance |
| Are counts meaningful? | [E11 report](results_v3/queries_report.json), [query code](scripts/benchmark_queries.py) | Warmup omitted, reference audit charged to the wrong side, or baseline recomputes already shared work |
| Does memory reduce actual error? | [E10 CSV](results_v3/innovation_raw.csv), [protocol](experiments_innovation.json) | Gains vanish against a strong matched-time/memory method |
| Is neural inference necessary? | E10 `constant_certificate` comparison | Current comparison couples the neural predictor and proposal; no isolated proof of neural necessity |
| Is this beyond known methods? | [Primary references](references.json), [manuscript](MANUSCRIPT.md) | Equivalent visibility-cache/certificate/control-variate interface already established |

## Evidence boundaries

E10 has twelve new motion scenes, three fixed sampling seeds and 18,144 frame-method records. Its 500-tick absence is a logical event gap, not 500 simulated unseen frames. Geometry updates are authoritative. Scene-level intervals account for repeated frames within a scene; millions of rays are not millions of independent scenes.

E11 has eight further scenes and five prescribed times, with three known appearance readouts at each scene-time. It fully charges the initial anchor visibility queries. The stronger baseline already shares visibility across readouts. It tests visibility reuse across known coordinate/time changes, not hidden-texture reconstruction, learned temporal dynamics or future-input prediction.

The main CPU timing compares small batches and excludes common feature/prior preparation and independent reference audits. Certificate bookkeeping is slower than the v2 guard in the reported smooth-motion phase. Query savings are not a measured frame-time or GPU gain.

## Proof and implementation gap

The sphere/segment certificate argument is in real arithmetic. The code uses float64 and a tolerance. Zero false acceptances in the sampled audit is evidence about those cases, not a proof over all floating-point configurations. Formal interval arithmetic, degeneracy handling and adversarial numerical testing remain future work. Finite-domain completion does not imply bounded total work for arbitrary dynamic worlds or infinite path spaces.

## Decisive next experiments

1. Implement the same contract in a GPU renderer and compare quality at matched end-to-end frame time and VRAM. Include certificate checking, memory traffic, updates and fallbacks.
2. Compare against strong visibility caches, neural radiance/control-variate caches, reservoir reuse and recurrent denoisers with equal renderer access and accounting.
3. Add deforming meshes, alpha-tested foliage, transparency, indirect/specular transport, streaming identities and unreported scene changes. Measure false-certificate rates and recovery.
4. Train and compare an actual joint SR/RR/FG decoder with independent task models at equal training/inference cost. Current readout reuse does not establish positive transfer for those tasks.
5. Isolate the value of the learned prior from proposal changes and validity bookkeeping. Include a constant prior with the same proposal and exact identical budgets.

Negative or null results should be retained. Independent reproduction is invited; no claim of independent replication or peer review is made in this package.

## How to report an issue

Record the immutable Hub commit, operating system, Python/NumPy versions, protocol and seed, command, expected versus observed value, and a minimal reproducer. Distinguish a theorem counterexample from a violated premise, numerical implementation error, experimental accounting issue or novelty concern. Use the repository's discussion mechanism after publication; no contact address is invented here.