docs/DATASET_GUIDE.md

Data guide

All rows are synthetic scientific artifact records. The dataset has no private participant data, real-world task traces, or trained model weights. Rows are fully disclosed and should not be treated as held-out evaluation examples.

The three tables are derived from the existing Revision 3 records. They do not add scientific experiments or increase the reported verification counts.

cases

• case_id: stable family name and zero-based index.
• family: named construction or verification family.
• theorem_reference: relevant numbered result in the unchanged Revision 3 manuscript.
• source_record, source_collection, source_index: exact provenance for the JSON record and collection element.
• request_count: reported request count, or a direct derivation from supplied center/value lists. Sentinel cases include the three added requests. -1 means not available.
• state_dimension, observed_dimension, variation_rank: recorded metadata; -1 means not recorded, not applicable, or mixed within the source case.
• rho: exact rational string when recorded, otherwise an empty string.
• orders_checked: reported finite enumeration count; -1 means not recorded. This is not an all-instance proof or a runtime benchmark.
• contains_float_diagnostics: whether the preserved case payload includes floating-point diagnostics. Exact algebraic decisions do not use those diagnostics.
• case_json: original complete case object serialized as JSON text. Rational values stay strings such as 1/2, avoiding binary floating-point conversion and irregular nested Arrow schemas.

Case families are balanced_half, balanced_general_rho, stochastic_and_spd, scalar_exact_hitting, scalar_additive, sentinel_recurrence, sentinel_partition, one_interruption, observable_rank, and observable_quotient.

claims

claim_id, status, statement, assumptions, limitations, supporting_code, and record_json normalize the ten entries in verification/claim_ledger.json. The three list columns contain strings. Empty lists mean no entry was provided in that particular ledger field, not that the theorem has no assumptions. Read the manuscript for complete hypotheses. record_json preserves the full ledger entry.

verification_suites

suite_id, source_record, verifier, scope, and arithmetic identify the audit. assertions_count contributes to the 711,899 total. placements_checked contributes to the separate 192,780 marker-placement count. A suite has zero in the inapplicable count field. details_json preserves non-list summary fields, including per-property counts where supplied.

Indices and exact arithmetic

Request and fragment indices in Python and exported records are zero-based. Always pair a fragment list with its own centers. The fluid verifier's symmetric tight example splits the positive-center label; the manuscript also illustrates the equivalent negative-center version. These are both valid, differently labeled schedules.

Use fractions.Fraction for rational strings:

from fractions import Fraction
import json

case = json.loads(row["case_json"])
centers = [Fraction(x) for x in case.get("centers", [])]

No train/test disjointness, distributional coverage, or model-learning benefit is inferred from these finite rows. The validation split name is a dataset packaging convention.