DATA_DICTIONARY.md

# Evidence and data dictionary

All scenes are procedural and synthetic. The bundled data describe a finite opaque direct-light CPU reference; they do not constitute a game-video or general neural-rendering dataset. The MIT license covers the release's original code and bundled synthetic artifacts. Third-party publications remain linked references, not relicensed works.

## E10: motion and sequential residual elimination

`results_v3/innovation_raw.csv` has 18,144 rows. The row key is `(scene, seed, phase, phase_frame, method)`. Twelve scene IDs are 400–411; three sampling seeds are 5101–5103. Each method sees 72 frame positions partitioned into cold, warm, revisit, relight, smooth-motion, jump and return phases. See the frozen `experiments_innovation.json` for all values.

| Column | Meaning / units |
|---|---|
| `scene`, `seed` | Procedural scene and sampling seed identifiers |
| `phase`, `phase_frame`, `frame` | Event phase, frame index within phase, overall frame index |
| `method` | Named estimator/ablation; `unsafe_eliminate` is a deliberate invalidation failure |
| `expected_mse` | Per-record diagnostic conditional expected mean-squared linear-RGB error under the estimator's sampling contract; see the risk calculation in code |
| `observed_mse` | Mean squared error of the actually sampled linear-RGB output against the reference |
| `rays` | Total online physical segment queries for this record; not rays per pixel |
| `receivers` | Number of queried surface receivers; use `rays / receivers` for queries per receiver |
| `runtime_seconds` | Measured estimator batch runtime; common preparation and separate reference audit are excluded |
| `memory_bytes` | Reported estimator memory footprint; not process RAM or GPU VRAM |
| `certified_fraction_before` | Fraction of relevant terms accepted as known before new queries |
| `false_certificates` | Number of accepted facts contradicted by the independent reference audit |
| `enclosure_violation` | Maximum diagnostic violation of the output enclosure for the record |
| `enclosure_mean_width` | Mean width of the direct-light output enclosure |
| `negative_channel_fraction` | Fraction of estimated color channels below zero before any display clamp |

Reports aggregate by scene/phase/method. Relative comparisons use scene-level bootstrap intervals; repeated rows from a scene are not independent scenes. `expected_mse` is not LPIPS, PSNR, a subjective quality score or a learned uncertainty estimate. Read `scripts/benchmark_innovation.py` and `aureole/innovation.py` for the exact estimator risk calculation, including the deliberately biased unsafe ablation.

## E11: spatial, known-time and appearance queries

`results_v3/queries_raw.csv` has 40 rows: scene IDs 600–607 and times 0, 0.25, 0.5, 0.75, 1. Each row supports three known appearance readouts. Anchor observations are frozen before queries. This is not a causal prediction benchmark for unknown future actions.

| Column | Meaning / units |
|---|---|
| `scene`, `time` | Procedural scene and prescribed normalized query time |
| `terms` | Number of finite receiver-light visibility terms queried at that time |
| `certified_terms` | Terms reused under valid spatial/motion conditions |
| `fresh_queries` | Physical queries required for the unresolved complement |
| `shared_fresh_baseline_queries` | Fresh visibility baseline that already shares work among all three readouts |
| `separate_readout_baseline_queries` | Weaker baseline that recomputes independently for each readout; not used for the headline |
| `false_certificates` | Accepted reused terms contradicted by full reference visibility |
| `max_linear_rgb_error` | Largest absolute linear-RGB difference among the tested readouts |
| `lookup_trace_seconds` | Measured lookup plus remaining physical-query time |
| `three_readout_seconds` | Measured time for the three appearance readouts |

Initialization queries are charged once per scene in `queries_report.json`, not repeated in every CSV row. The headline calculation is `1 - (147456 + 535303) / 2949120 = 0.7684872097`. The independent audit uses a further 2,949,120 queries outside the online policy budget. Neither number is an end-to-end speedup.

## Arrays, prior and provenance

- `.npz` artifacts use NumPy array archives. Load with `numpy.load(path, allow_pickle=False)` and inspect `.files` for keys; no serialized executable model object is required.
- `models/visibility_prior.npz` is the frozen v2 prior. Its training/test partitions and metrics are in `results/training.json`; its role is documented in `MODEL_CARD.md`.
- Frozen configuration hashes are embedded in the two v3 reports. File-level hashes are in `CHECKSUMS.sha256`.
- `results/` contains earlier evidence; `results_v3/` contains current scientific evidence; `results_publication/` contains packaging checks only.
- No additional personal or externally licensed game assets are required to generate these synthetic scenes. Generalization to realistic content remains open.