# Dataset schema and provenance

The dataset contains five separate synthetic configurations, with 240 rows total. Each has a single complete `train` split for loading compatibility, rather than a supervised-learning evaluation split. Files are declared explicitly in the root dataset card; auxiliary CSV, JSON, PDF, code, and archive files are not additional loading inputs.

## Configurations

| Configuration | Rows | Fields |
|---|---:|---|
| `six_wing_closure` | 64 | `mask`, `active_wings`, `active_names`, `rank`, `formula`, `source_bits`, `retained_fraction`, `closure_rounds` |
| `photon_frontiers` | 44 | `modes`, `messages`, `logical_bits`, `storage_cutoff`, `symmetric_cutoff`, `storage_mean_photons`, `symmetric_mean_photons`, `cutoff_ratio` |
| `covariance_plateau` | 54 | `messages`, `modes`, `cutoff`, `physical_dimension`, `matched_dimension`, `minimum_normalized_group_defect` |
| `approximate_query_energy` | 72 | `independent_query_bits`, `modes`, `query_error`, `necessary_mean_photons` |
| `six_label_encoding` | 6 | `basis_index`, `sector_photon_number`, `lily`, `tachy`, `raven`, `enya`, `evie`, `kaya` |

## Interpretation

`six_wing_closure` enumerates all six-generator bitmasks from 0 to 63. `active_names` is semicolon-separated and empty when no generator is active; readers may infer that empty value as null. `rank` is the computed retained task dimension; `formula` is the comparison prediction. `retained_fraction` divides the retained rank by 36.

`photon_frontiers` compares register-only photon resources over the stated mode/message grids. `impossible` denotes failure of the requested coherent contract in that setting; `undefined` denotes an unavailable ratio, including division by a zero storage cutoff. Sentinel-bearing columns require explicit conversion for numerical use and may load as strings.

`covariance_plateau` holds the 64-label, two-mode example at cutoffs 10 through 63. The defect follows the manuscript's phase-registered group-average metric and is not an experimental error probability.

`approximate_query_energy` is a lower-bound table over declared bit counts, mode counts, and query errors. Its `necessary_mean_photons` values are necessary resource bounds, not achieved device operating points.

`six_label_encoding` is a headered representation of the original real 6×6 encoding matrix. Rows are physical basis coordinates; columns are the named logical labels. `basis_index=0` is the vacuum coordinate, with `sector_photon_number=0`; indices 1 through 5 span the four-photon sector, with `sector_photon_number=4`. Those indices label a basis and do not assert a fabricated gate layout. The six amplitude columns reproduce the original numeric tokens exactly.

## Generation and boundaries

The first four loading files come directly from `code/reproduce.py`. That script also produces the original headerless matrix `data/six_label_encoding.csv`. Dataset edition v1.0.2 adds headers and the two basis metadata columns in `dataset/six_label_encoding.csv`; the original file is preserved and remains the reproduction target.

These are deterministic finite constructions and model-based resource calculations. There is no population sampling, annotation workforce, training corpus, measured optical trace, or personal data. The grids and noiseless assumptions are deliberate and narrow. The tables are suitable for inspecting and reproducing the given examples; they do not establish benchmark superiority or realistic device performance.

The manuscript defines the mathematical quantities. `requirements-tested.txt` and `verification/hf_release/` record the software environment and fresh checks. The dataset edition's original source files remain byte-identical to core v1.0.0.