# Verification dataset

The `channel_comparisons` configuration contains **1,247 saved synthetic channel comparisons**, exposed as a single `validation` split. This split name identifies mathematical verification records, not a machine-learning train/validation partition. Cases include endpoints and curated degeneracies as well as seeded random examples; no claim of independent identically distributed sampling is made.

| Family | Rows | Original report |
|---|---:|---|
| `deficiency` | 556 | `research/data/deficiency.json` |
| `purification` | 375 | `research/data/purification.json` |
| `orbit_audit` | 96 | `research/data/orbit_audit.json` |
| `catalogue` | 220 | `research/data/catalogue.json` |

| Field | Type | Meaning |
|---|---|---|
| `case_id` | string | Stable family plus zero-padded source record index. |
| `family` | string | One of the four modules above. |
| `category` | string | Source category, or an explicit family-level fallback. |
| `q` | integer | Number of hidden hypotheses in the compared experiment. |
| `analytic_deficiency` | float | Formula prediction for directed worst-state TV deficiency. |
| `unrestricted_lp_deficiency` | float | Independently optimized unrestricted stochastic-kernel LP value. |
| `absolute_error` | float | Absolute difference between those two values. |
| `constructed_kernel_tv` | float or null | Attained TV for the saved constructive kernel when that report stores it; null is not zero. |
| `comparison_tolerance` | float | Common export check threshold, `1e-9`; individual modules retain their original test tolerances. |
| `passed` | boolean | Formula/LP discrepancy and any saved kernel discrepancy meet this export tolerance. |
| `source_report` | string | Relative path to the detailed original JSON. |
| `source_record_index` | integer | Zero-based offset inside that report's `records` array. |
| `parameters_json` | string | Original non-result fields, serialized as JSON to keep one stable tabular schema. |

`verification/summary.json` gives the counts and largest discrepancy. All detailed reports, including separate Gaussian, finite-field and physical-flow checks, remain in `research/data/`. They have different schemas and are intentionally excluded from the default dataset configuration.

## Regenerate

After intentionally rerunning the research suite:

```bash
python scripts/build_verification_dataset.py
python scripts/update_manifest.py
```

The normalization performs no new experiments and does not alter the scientific results. Every normalized row links back to the saved source record. Floating-point agreement supports an implementation check, not a proof, an outward-rounded interval guarantee or a physical measurement.