Datasets
Bar datasets
Built by research/dataset.py + model_generation/ with quality gates
(coverage, ordering, duplicates, horizon boundaries). Each dataset is a
versioned artifact with a fingerprint (ds_* id) used in manifests and run
snapshots.
Tick datasets (CHG-0041 / TICK_DATASET_META v2)
research/mt5_tick_dataset.py acquires ticks from the **certified adapter
surface** (copy_ticks_range, COPY_TICKS_ALL), with:
- bounded, deduplicated acquisition (post BUG-188 semantics),
- immutable parquet store with content fingerprints,
- provenance metadata v2 (symbol, window, source, acquisition identity),
- cache-hit idempotence — re-running acquisition never duplicates.
After acquisition the dataset is fully offline: research needs no broker.
Rules
- A dataset is immutable once fingerprinted — corrections produce a new
dataset, never a mutation.
- Every experiment names its dataset ID; manifests and run snapshots carry it.
- Dataset-fairness: A/B/C model comparisons must consume equal data windows,
seeds and splits (the TASK-04 protocol).
- Provenance fields may be
NOT_RECORDED(honest) — they are never invented.