[0z5a]

The original RLTT data converter reads problem/answer; passing the earlier public prompt/label JSONL gives an empty question and gold label. Add a separate public-input profile carrying both interfaces, with the pinned source instruction and default chat options. All 12,498 rows produce exactly matching original-converter and native prompts, token IDs and labels. The earlier data remains unchanged.

Incremental draft on #16. No production training, rollout or reward code changes.

Comparison	Original / parent	Candidate	Speedup
RLTT conversion of public inputs	v1 row gives empty question and gold	12,498 complete rows match VIME	Not a timing comparison
Complete original RLTT vs ScaleRLT	Not run	Not run	Not measured
GPU peak memory / held-out time-to-quality	Not run	Not run	Not measured

Validation: 62 existing CPU data/reward/evaluation tests pass without skips. Two fresh preparations produce byte-identical five JSONL files, five Parquet files, manifest and token rosters. The 1,024-token source/default-chat training profile keeps 7,479 rows with 19 explicit exclusions; all 256 development and 500 held-out rows remain. Official Ouro AutoConfig loading succeeds in the existing environment; complete VIME/MCore parser import fails on missing local Triton, with both receipts retained.

The unchanged original configuration functions and declared verl 0.6.1 source also expose a batch-accounting difference: conditional on DP4 and no private changes, CLI 32/8 produces 2,048 completions per rollout, not the nominal 256. Counts and budget consequences are explicitly source-derived; no full original dataloader/FSDP/optimizer run is claimed. Grader, optimizer, private model and actual training counts remain separate baseline qualifications.

See docs/rltt_source_contract.md for the complete contract, speed-comparison table, source pins, reproduction and limitations; raw results and hashes are under benchmarks/results/rltt-source-contract/.