curate-trajectories

To prevent trajectory data leakage, you can curate trajectories into a provenance-rich dataset that cross-checks against holdout id sets and hard-blocks guarded pools. This ensures only train-safe rows are emitted for downstream training.

A decontaminated trajectory pool is a train-safe dataset generated by indexing trajectories, tagging splits from frozen evidence, and applying contamination checks. The output includes a selection hash, corpus hash, and an auditable manifest for verification.

You track dataset provenance by building a local index with full provenance fields and corpus hash metadata for each trajectory. This index records the origin and split tags, enabling hash-stamped selections for downstream audits.

Yes, you can map trajectories to train, dev, holdout, or none by referencing a frozen splits.json file. The skill annotates index records with these split tags to create guarded pools for safe training and distillation.

The best way to audit trajectory selections is to resolve filtered subsets into a named selection and compute a selection hash. This process emits a manifest containing the splits_sha256 and corpus hash for full downstream verification.

No, this trajectory curation workflow operates without external dependencies. It independently indexes your local trajectory files, applies contamination checks, and emits the decontaminated pool directly to your environment.