[0z5a]

Prepare immutable MATH inputs for the pending native recurrent RL quality experiments. The corpus now has verified source hashes, explicit answer repairs and exclusions, disjoint development/held-out partitions, and an official Ouro-Thinking tokenizer audit. This is a reconstructed public-data contract, not an exact reproduction of RLTT’s private JSONL.

Incremental draft on #14. No training runtime or reward implementation changes.

Input / check	Result
Prepared train / development / MATH-500 / reserve	7,498 / 256 / 500 / 4,244
Explicit Ouro thinking training profile, cap 1,024	7,480; 18 over-cap rows recorded
MATH-500 questions or labels removed	0
Actual Dataset and canonical/missing-answer reward checks	All 12,498 rows pass each check
Official-tokenizer row/profile observations	37,494
Unit, reward and evaluation tests	62 passed

Speed comparison	Parent #14	Candidate	Speedup
Audited quality inputs	No corpus/length roster	Reproducible inputs and full receipts	Not a timing measurement
Full online RL	Pending	Pending	Not measured

Regeneration is byte-identical. Source integrity, the 4 solution-hash-bound label repairs, 2 excluded train rows, 2 retained unknown difficulty labels, and all prompt length exclusions are recorded. Initial preparation/test failures are retained. Known-answer reward routing is data plumbing, not model accuracy or convergence; no GPU/model execution is claimed.

Full report: docs/math_quality_inputs.md. Raw hashes, row rosters, token-ID hashes and test receipts: benchmarks/results/math-quality-inputs/.