# Dataset schema and provenance

This is a small, document-derived **theoretical research corpus**, not an
empirical benchmark, a record of an implemented RSI system, or a collection
of machine-checked proofs. Its two JSONL views preserve source material and
make it accessible to ordinary dataset tooling.

## Configurations

| Configuration | JSONL file | Records | Content |
|---|---|---:|---|
| `manuscript` (default) | `data/manuscript_sections.jsonl` | 19 | Lossless, ordered LaTeX source chunks |
| `claims` | `data/claims.jsonl` | 16 | Complete claim objects from the supplied claim register |

Each configuration uses the single split name `train` as a Hugging Face
loading container. That name does not establish training data suitability,
a train/test separation, a held-out evaluation set, or any experiment.
The two configurations overlap substantially: the claims describe results
and limitations already discussed in the manuscript. They must not be
treated as independent evidence or mutually held-out data.

## Common fields

Every row in both JSONL files has exactly these nine string-valued fields.

| Field | Meaning |
|---|---|
| `id` | Unique within the complete release: `manuscript-000` through `manuscript-018`, or `claim-` plus the source claim identifier. |
| `title` | The source section heading, a descriptive label for front/end matter, or the complete source claim statement. |
| `content` | Verbatim LaTeX source fragment, or a JSON serialization of the complete original claim object. |
| `content_format` | `latex` for manuscript chunks; `json` for claim objects. |
| `source_path` | Repository-relative authoritative source: `manuscript/Iyeumiru_R_Rayo_RSI.tex` or `metadata/claims.json`. |
| `source_locator` | A JSON-encoded location object for LaTeX, or an RFC 6901-style JSON Pointer such as `/claims/0`. |
| `record_type` | `manuscript_source_chunk` or `claim_register_entry`. |
| `evidence_status` | Manuscript-wide source status, or the exact `status` string in the original claim object. This is reported source metadata, not an independent verification judgment. |
| `source_sha256` | SHA-256 of the entire authoritative source file, not of the individual fragment. |

All JSON files and JSONL files are UTF-8. Each JSONL row is one JSON object
on one physical line; newlines in `content` are JSON escapes. No remote code
or custom dataset loader is required.

## Manuscript boundaries and exact reconstruction

The 19 manuscript rows contain, in source order:

1. Preamble, title matter, author credit, and research status.
2. Abstract, keywords, and table-of-contents commands.
3. Ten main sections.
4. Five appendices.
5. Research conclusion.
6. References and the closing document command.

The first appendix includes the `\appendix` switch. Blank lines and all
LaTeX commands remain in their original positions. The chunks are not
standalone compilable documents: they depend on the preamble and definitions
in the complete authoritative source.

A manuscript `source_locator` is itself a string containing JSON, for example:

```json
{"line_start":1,"line_end_inclusive":64,"utf8_byte_start":0,"utf8_byte_end_exclusive":2705}
```

Line numbers are one-based and inclusive. Byte ranges are zero-based UTF-8
offsets with an exclusive end. Byte offsets remain exact in the presence of
the manuscript's Unicode characters.

Concatenating the `content` fields in their existing JSONL order and encoding
the result as UTF-8 reconstructs **all 100,396 original LaTeX bytes**, with
SHA-256:

```text
af507774f200e005535ee5f700c3f1ef917bc60e0744d460b356c5f0afcc4922
```

## Claim payloads and evidence labels

Each claim row contains the complete corresponding object in
`metadata/claims.json`, including its source identifier, statement, status,
assumptions where supplied, source label where supplied, and limitations.
Parsing its `content` as JSON yields an object equal to the original object.
Formatting and key order are normalized; object values are not rewritten.

The original registry's manuscript-wide metadata remains in
`metadata/claims.json`; it is not duplicated into an extra claim row. That
source file has SHA-256:

```text
0e22ea2ff43f482f2550aa343ecc0f00e8b3e160a83d5bbf561e6c8c7518bcd4
```

| Claim evidence status | Rows | Interpretation in the source |
|---|---:|---|
| `proof_supplied_in_manuscript` | 12 | A written proof is supplied, with the stated scope and assumptions. |
| `derivation_supplied_in_manuscript` | 1 | A written derivation is supplied. |
| `required_not_implemented` | 1 | An implementation obligation, not an achieved implementation. |
| `open_not_demonstrated` | 1 | An open research hypothesis. |
| `not_claimed_not_established` | 1 | A broad claim explicitly not asserted or established. |

The manuscript rows use
`theoretical_manuscript_not_independently_verified`. Neither that label nor
the claim labels assert mechanized verification, empirical acceleration,
independent review, reproduction, or established literature priority.
Hypothetical numerical examples remain hypothetical source content.

## Rebuilding and validating

From the repository directory, using Python 3.9 or newer:

```bash
python tools/build_corpus.py --check
```

This standard-library-only command checks byte-for-byte agreement between
the supplied JSONL views and deterministic conversions of the authoritative
sources. It also checks common schemas, unique identifiers, exact manuscript
reconstruction, exact source byte locators, and equality of every parsed
claim payload.

To regenerate the JSONL views after an intentional source edit:

```bash
python tools/build_corpus.py
```

Regeneration changes release bytes when source content changes. Update the
release checksum manifest and provenance before uploading an edited release.
Passing these conversion checks validates packaging fidelity; it does not
validate the manuscript's mathematical claims or demonstrate the proposed RSI
architecture.

## Source and rights

The PDF, LaTeX, original README, and claim register came from the supplied
`Iyeumiru_R_Complete_Research.zip`. The four hashes in that archive's
`SHA256SUMS.txt` were checked and matched before conversion. The PDF and
LaTeX remain authoritative for the full manuscript.

The supplied source package contains no explicit license notice. Dataset
conversion does not introduce a new license or grant rights absent from the
source. Preserve the author credit and consult the repository's rights note
before redistribution or reuse.