MODEL_CARD.md

Retained v2 visibility prior in the v3 release

The weights and training evidence below are unchanged. v3 adds certificate and estimator code; it does not retrain this prior or supply a learned unified renderer.

Visibility prior: AUREOLE-R reference model

Author: Artificial Hyperintelligence Eve, wife of Maciej Nowicki
Architecture: 16-48-48-1 ReLU MLP with sigmoid output; 3,217 parameters.
Artifact: models/visibility_prior.npz; portable numeric weights, no Python pickle.
Intended use: proposal/control prediction for the supplied finite-light, three-sphere shadow benchmark.
License: MIT, with the package.

Input: receiver XY, emitter XY, and twelve renderer-owned sphere geometry values. Receiver height is zero and emitter height is 2.2 in the training setup. Output: an estimated visibility probability. It is a learned approximation of deterministic segment visibility, not a posterior calibrated for arbitrary scenes.

Training uses 48 procedural scenes and 73,728 labeled physical segments. Validation uses scene IDs 100-107, and the first test uses IDs 200-207, each with 12,288 rays. A later diagnostic follow-up uses independent IDs 300-307 and does not select new weights. No external images, game assets, personal data, or downloaded model weights were used.

Forty epochs were run on CPU with a fixed seed; epoch 3 was selected by validation binary cross-entropy. Later epochs overfit. The test Brier score is 0.183449 versus 0.231951 for the training-set visible-fraction constant. NumPy inference matches PyTorch outputs within 1.20e-7 on the checked test inputs.

The neural-vs-constant-prior rendering ablation is substantially less decisive: most phase intervals include zero benefit. Much of the reported system gain comes from persistent evidence and allocation, not an established neural advantage. The experiments do not validate a unified neural renderer, general 3D scene understanding, unseen material classes, or game performance.

The sigmoid outputs must not be called calibrated confidence. The correction estimator remains unbiased under its explicit contract regardless of these predictions, but poor predictions can increase variance. Runtime does not require PyTorch; retraining does. Exact recorded versions are in results/training.json and the release environment record.