curate-trajectories

What problem does it solve?

Trajectories pile up as loose per-task JSON across runs, and once used for training, distillation, or RL they can leak frozen developer holdout. This skill provides a provenance-tracked, contamination-safe workflow to import, tag splits, perform hash-stamped selections, and emit a decontaminated pool that preserves holdout integrity.

Core Features & Use Cases

• Index + attach provenance: Build a local index (one record per trajectory) at .understudy/curate-trajectories/index.jsonl with full provenance fields and corpus/hash metadata.
• Tag splits from capture-evidence: Map trajectories to train/dev/holdout/none using the frozen splits.json and annotate index records for guarded pools.
• Query as a hash-stamped selection: Express subsets by provenance filters, resolve to a named selection, compute a selection hash, and emit a manifest for downstream audits.
• Contamination check & emission: Cross-check against holdout/dev id sets, produce a contamination report, and hard-block guarded pools unless overridden.
• Emit decontaminated pool: Produce train-safe/distill-safe pools along with the selection hash, splits_sha256, corpus hash, and row counts for auditability.

Quick Start

Index your trajectories and generate a train-safe, decontaminated pool using the built-in workflow.