Blame for docs/DATASETS.md

# Dataset configurations and provenance

This Hub edition exposes selected existing metadata and synthetic numerical records as UTF-8 JSON Lines. `tools/build_hf_data.py` deterministically rebuilds the four files using the preserved research reports; it does not rerun experiments or derive new research conclusions.

Each configuration has one split named `train` to support standard Hub loading. This label does not describe a train/test experiment. Counts: 15 + 63 + 299 + 141 = 518 rows. All rows within a configuration use the same keys.

## `claims` — 15 rows

Source: the `claims` array of `research/docs/claims.json`.

| Column | Type | Meaning |
|---|---|---|
| `id` | string | Original claim identifier C01–C15 |
| `title` | string | Original claim title |
| `contract` | string | Assumptions and task definition |
| `proof_status` | string | Written-proof and internal-review status |
| `novelty_status` | string | Original historical-priority assessment |
| `source` | string | Proof-source path relative to repository root |
| `labels` | list of strings | LaTeX statement labels |
| `theorem_numbers` | list of strings | Statement numbers, including non-theorem statements |
| `printed_pages` | list of integers | Printed manuscript pages, not PDF page indices |
| `evidence_files` | list of strings | Repository-relative evidence paths; may be empty |

Normalization prefixes `research/` to original research-relative paths and converts printed page numbers to integers. The three location lists have matching order and length. Textual claim assessments are preserved exactly.

## `announcement_family` — 63 rows

Source: `research/results/announcement_family.json`, key `family`.

| Column | Type | Meaning |
|---|---|---|
| `q` | integer | Hidden dimension, 3 through 65 |
| `r` | integer | `q-1`; omitted dimension and permitted announcement-cell count in this family |
| `p` | integer | Stored archive rank, always one |
| `k` | integer | Permitted announcement-cell count, equal to `r` |
| `t_star` | float | Optimal total squared archive weight on singleton actions |
| `exact_cost_times_B` | float | Floating-point evaluation of the proved constant `c_r` |
| `max_singleton_cost_times_B` | float | Largest constructed singleton cost times budget |
| `pair_cost_times_B` | float | Constructed unresolved-pair cost times budget |
| `best_block_cost_times_B` | float | Block-restricted optimum, one |
| `general_lower_envelope_times_B` | float | General lower bound from the larger model |
| `strict_block_improvement_times_B` | float | Difference between block optimum and constructed global optimum |
| `pair_precision_eigenvalues_over_B` | list of floats | Two positive pair-sensor precision eigenvalues divided by budget |

The word `exact` in a column name refers to the underlying theorem formula. Values in these JSONL rows are binary floating-point evaluations. Exact rational identities for the three-coordinate example are retained separately in the source report under `exact_rational_certificate`.

## `memory_frontier` — 299 rows

Source: `research/results/codesign_results.json`, key `memory_frontier`.

| Column | Type | Meaning |
|---|---|---|
| `q` | integer | Hidden state dimension |
| `p` | integer | Archive rank |
| `r` | integer | Omitted dimension `q-p` |
| `B` | float | Per-selected-sensor squared-gain budget |
| `shared_optimum` | float | Formula `(q-p)^2/(qB)` |
| `announced_optimum` | float | Formula `(q-p)/(qB)` |
| `coordinate_storage_shared` | float | Shared-sensor cost of the compared coordinate archive |
| `shared_constructed` | float | Cost evaluated on the constructive shared-sensor design |
| `max_random_dual` | float | Largest dual value found in the finite random audit; not an independently certified optimum |

## `divisible_blocks` — 141 rows

Source: `research/results/codesign_results.json`, key `divisible_blocks`.

| Column | Type | Meaning |
|---|---|---|
| `q` | integer | Hidden state dimension |
| `r` | integer | Omitted dimension; archive rank is `q-r` |
| `k` | integer | Number of announcement blocks in the tested divisible construction |
| `B` | integer | Squared-gain budget |
| `optimum` | float | Applicable proved optimum in the stated integer family |
| `constructed` | float | Floating-point evaluation of the attaining block construction |

## Selection and intended use

The four files expose a claim index and three coherent parameter families that are useful for browsing and comparing designs. They are a subset of the full research evidence. Repair cases, quantum diagnostics, certificate-menu results, exact certificates, figure data, and seeds remain available under `research/`.

No new observations, trainable model, external benchmark, or experimental measurements were created for this packaging step. The data reflect the chosen finite models, parameter ranges, tolerances, and synthetic constructions. They do not measure practical AI memory savings, energy efficiency, or hardware performance.

For local loading, install `datasets` in an environment of your choice and call:

```python
from datasets import load_dataset
rows = load_dataset("json", data_files="data/announcement_family.jsonl", split="train")
```

Source documentation for this Hub layout: [Dataset cards](https://huggingface.co/docs/hub/datasets-cards) and [manual configuration](https://huggingface.co/docs/hub/datasets-manual-configuration), consulted 20 October 2026. All four configurations explicitly select their own file; unrelated research JSON files are not automatically combined into a dataset.