DATASET_GUIDE.md

Data and retrieval guide

This dataset is a research release. It packages mathematical text, source code, a claim ledger, transparent review tasks, exact rational objects, and finite experiment records. It contains no human-subject observations, trained weights, or private interaction logs. The supplied attribution, dedication, and image remain original research-package content. Reuse rights are described in RIGHTS.md.

Independent configurations

The README YAML names each subset explicitly so unrelated schemas are not accidentally merged. All eight viewer configurations use JSONL, avoiding cross-configuration builder inference between CSV and JSON. Original CSV tables remain unchanged in results/. Exact record counts are in data/counts.json.

Configuration	Split	File	Role
research_corpus (default)	corpus	data/research_corpus.jsonl	Source-grounded text and code chunks
claims	ledger	data/claims.jsonl	Lossless claim/evidence/boundary columns from CLAIMS.md
review_tasks	tasks	data/review_tasks.jsonl	Eight transparent expert/agent review tasks
exact_family	observations	data/exact_family.jsonl	Five exact family sizes, dimensions, positivity and spectra checks
generic_compiler	observations	data/generic_compiler.jsonl	Thirty generated instrument/compiler cases
snapshot_frontier	observations	data/snapshot_frontier.jsonl	Six sharp factor-two fixtures
long_horizon	observations	data/long_horizon.jsonl	128 finite histories of 256 supplied labels
rewrite_approximation	observations	data/rewrite_approximation.jsonl	72 perturbation histories

Split names organize artifact types. They do not imply statistically independent training, validation, or test data. The review questions and source answers are public together. This is not a held-out benchmark, and results on those questions do not establish general agent capability.

Load locally

Install requirements-data.txt separately from the mathematical environment if you only need retrieval. From the repository root:

from datasets import load_dataset

corpus = load_dataset("json", data_files={"corpus": "data/research_corpus.jsonl"}, split="corpus")
claims = load_dataset("json", data_files={"ledger": "data/claims.jsonl"}, split="ledger")
traces = load_dataset("csv", data_files={"observations": "results/long_horizon.csv"}, split="observations")
print(corpus[0]["source_path"], corpus[0]["text"][:100])

After publication, the eight configurations can be loaded by repository ID and configuration name as shown in the README. No trust_remote_code argument is required. Use snapshot_download(..., repo_type="dataset", revision="<actual commit hash>") to fetch the full manuscript/code/certificate tree at a fixed version. A named configuration loads only its selected tabular data; it does not install research code or supply all evidence files.

Corpus record schema

Every line of data/research_corpus.jsonl is an object with these fields:

Field	Type	Meaning
id	string	Deterministic source-path and zero-based chunk index
source_path	string	Repository-relative authoritative source
source_sha256	string	SHA-256 of the complete source bytes
format	string	latex, markdown, or python
evidence_level	string	Role of the text, such as written argument, implementation, or navigation
chunk_index	integer	Zero-based index within that source
char_start, char_end	integer	Half-open Unicode character offsets in the UTF-8-decoded source
text	string	Exact source slice at those character offsets

tools/build_agent_data.py deterministically exports eleven selected text/code sources. It splits at paragraph boundaries when possible, with a maximum of 12,000 characters per chunk. It does not summarize the manuscript, translate equations, infer proof validity, or manufacture theorem text. A theorem can cross a chunk boundary: load neighboring chunks or the complete source for proof review. PDF pages, images, and large raw JSON certificates are not embedded in the corpus; their canonical files are available separately.

To check a chunk, hash the complete source file and compare source_sha256, decode it as UTF-8, and compare source_text[char_start:char_end] with text. python verify_release.py performs these checks. Character offsets count decoded characters, not bytes or model tokens. Source hashes make cross-version drift detectable.

evidence_level classifies the document's role. A written_mathematical_argument value means the text is an argument in the supplied manuscript, not that a theorem prover or independent expert certified it. Source comments, documents, and retrieved chunks should be treated as research data rather than higher-priority operational instructions.

Claims and tasks

data/claims.jsonl preserves the source ledger columns as claim, evidence, and boundary, with an ID, source path/hash, and a coarse status. not_established mirrors the source's explicit status; other rows are source_claim_with_stated_evidence. The latter is not a new endorsement or uniform validation level.

data/review_tasks.jsonl has id, task, required_sources, acceptance_criteria, task_type, and benchmark_status. For a flat interoperable table, required_sources is a JSON-encoded string array; parse it with json.loads. Acceptance criteria are review guidance, not hidden gold labels. All three JSONL formats have JSON Schemas in schemas/.

Experiment table dictionary

Fields use the source generator's names. Coordinate and rank fields count real/exact algebraic resources; they do not count bits, transient workspace, model storage, or acquisition cost.

Each table's JSONL export is derived deterministically from the same-named original CSV: integer cells become JSON integers, True/False become booleans, and decimal/scientific cells become numbers. The verifier compares every exported row and field with the parsed original CSV. Original textual representations and exact rational certificates remain in the source files.

Table	Key fields and interpretation
exact_family	n: latent states; exterior_basis: chosen orbit rank; nonlinear_continuous_scalars, affine_snapshot_scalars, affine_recurrent_scalars, rewritten_continuous_scalars: distinct resource contracts; entry/probability minima are numerical displays of exact family properties; spectrum flags refer to the supplied perturbation; elapsed_seconds is original wall-clock timing
generic_compiler	n, trial: fixture; exterior_rank: finite closure rank; signature_jacobian_rank: exact evaluated fiber/Jacobian rank; snapshot_scalars: affine snapshot dimension; basis_word_depth: maximum basis provenance length; extra_words and word_length: additional membership probes; max_LP_representative_update_error: finite floating-point diagnostic
snapshot_frontier	r, n: fixture parameters; exact_continuous_scalars, affine_snapshot_scalars, verified_jacobian_rank: sharp-example checks
long_horizon	trial, n, length: history; exact_error: numerical difference between full and compact reference trajectories; quantized_error: rounded implementation difference; uniform_bound: the analyzed rounding bound; step: quantization grid
rewrite_approximation	delta, trial: perturbation fixture; max_readout_error: observed finite-history maximum; uniform_bound: 8*delta/3

exact_error is the generator's field name; its CSV values are floating-point measurements, not exact arithmetic certificates. The all-horizon statements rely on the written proofs, not extrapolation from 256-step histories.

Exact objects and raw traces

results/exact_certificates.json stores the small exact family certificates. results/generic_certificates.json stores the thirty generic model/signature objects. Matrix entries are rational strings. Reconstruct with sympy.Rational; do not use eval or first convert to binary floating-point for an exact audit. See check_exported.py for the supported exported-family checker.

results/long_horizon_raw.json stores every initial distribution and supplied label sequence, plus the first twelve output steps. It does not store every intermediate output. results/hidden_reconstruction.json stores exact recovery observations and results at three sizes. The fixed-seed generator can reconstruct the remaining finite outputs.

Updating a release

After reviewing intentional source or documentation edits, run python tools/build_agent_data.py, then python tools/build_release_manifest.py, then python verify_release.py. Preserve the original research inventory. Changes to original scientific files need their own versioned research release and provenance treatment; do not simply replace the original inventory hashes to conceal changes. Update fresh-check records when rerunning validation. Checksums provide integrity checks, not signatures or trust guarantees.

The README configuration syntax follows the official Hugging Face data-file configuration documentation, checked 23 October 2026. Live Hub rendering is a post-upload check and is distinct from the local loader checks recorded here.