docs/DATA_CARD.md

# Synthetic triadic protocol table

`data/triadic_protocols.csv` contains all six permutations of A, B, C in lexicographic order. It is generated from the original `results/sharp_triad.json`, which is regenerated by the supplied verification suite.

Each row uses three markers and cyclic records $r_{AB},r_{BC},r_{CA}$, starting with all six cells present. The observer returns one when at least two cyclic records survive. Actions delete markers and guarded records; the fourth step is a common reset.

| Field | Type | Meaning |
|---|---|---|
| `order` | string | Complete three-action order |
| `permutation_sign` | integer | +1 for even, -1 for odd relative to A,B,C |
| `initial_cells` | integer | Allocated cells present initially: 6 |
| `observation_0` | integer | Initial chosen readout |
| `observation_1` through `observation_3` | integer | Readout after each primitive action |
| `observation_4_reset` | integer | Readout after the extra common reset |
| `extinction_step` | integer | First zero readout; 3 for odd, 4 for even |

There is no sampling, training corpus, personal data, annotation by human subjects, or empirical physical measurement. All rows are deterministic finite calculations. The six-row `test` split is a complete toy protocol table, not an unseen generalization benchmark.

Reproduce with `python tools/reproduce.py --full --output runs/data-check` and compare the regenerated `sharp_triad.json` with the CSV fields. The public card declares this one CSV configuration so unrelated result tables are not automatically combined into a dataset.

Version, source, and rights: research v1.0.0; packaging v1.0.1; source package preserved under `provenance/`. No source license was specified. The current release does not add a reuse license.