Curate Trajectories

Curate trajectories

Trajectories pile up as loose per-task JSON across runs (each row roughly {id, name, score, passed, assertion_results, steps, messages, end_state, finish_reasons, model, toolset, domain, seed, input/output_tokens, cost}). The moment anyone feeds them to training, distillation, or RL they stop being logs and become a dataset — and a dataset with no split hygiene silently kills heldout claims. This worker owns the trajectory dataset as a first-class, queryable, provenance-tracked, contamination-safe artifact: import → tag splits → select (hash-stamped) → contamination check → emit a decontaminated pool.

The store is provider-agnostic. The source is Lilac or a local JSON corpus; this skill owns the hygiene and provenance layer on top, not the browser.

Safety Gates

• Block by default. Any selection destined for a train / RL / distill / SFT pool is built excluding frozen dev+holdout. Including holdout (or dev) in such a pool is refused unless the developer passes an explicit override, and every override is logged in the manifest with who/when/why. Never silently include. A blocked selection is the safe outcome, not an error to route around.
• The frozen splits are the source of truth. Split membership comes from splits.json produced by capture-evidence, never re-derived by re-seeding or re-hashing here. If splits.json is missing or stale, stop and route to capture-evidence — do not guess membership.
• Local-first, no exfiltration. Index, query, and manifest stay under .understudy/. Never upload the corpus, never print message bodies, secrets, or raw payloads — operate on ids, provenance fields, counts, and hashes.
• Reproducible or it didn't happen. Every selection is named and hash-stamped (selection hash over the sorted row ids + filter expr + corpus hash). Downstream consumers cite the hash, not a manual grep.

Decision Gate

Use this skill whenever trajectories feed a train / RL / distill / SFT pool, or whenever you would otherwise hand-filter rows in bash (grep toolset=api, "drop the seed-7 rows", "keep only passes"). Hand-filtering is the #1 way a heldout claim silently dies; replace it with a hash-stamped selection. If the trajectories are only being eyeballed (no downstream training/claim), a manifest is optional — but the moment a number leaves the building, curate first.

This skill selects on provenance and contamination (split membership, dedup, outcome conflicts) — it does not judge whether a trajectory is correct or learnable. That pedagogical filtering (the useful-quadrant / surprise-gap cut) lives in local-distillation-lab's pedagogical arm, which consumes this skill's decontaminated pool.

Flow

1. Index + attach provenance. Build a queryable local index (one record per trajectory) at .understudy/curate-trajectories/index.jsonl from the run JSONs (and/or a Lilac export). Carry full provenance: task_id, model, toolset, domain, source_run_id, timestamp, outcome, split (split left unknown until step 2; back-fill model/toolset/domain from run-level meta when rows omit them). Record a corpus hash and per-source counts. The task id is the stable name (real exports put it in name, not the enumeration id). Don't require a per-row RNG seed — many workloads have none (the initial state is the seed); contamination safety keys on task_id↔split. Flag rows missing the fields needed to tag or filter, and keep them out of guarded pools.
2. Tag splits from capture-evidence. Load the frozen .understudy/capture-evidence/splits.json (its splits_sha256 goes in the manifest). Map each trajectory's task id to train / dev / holdout / none by the frozen split membership (the rows/row_ids lists). Tag every index record. Rows in no frozen split are none and are quarantined from guarded pools unless explicitly admitted.
3. Query as a hash-stamped selection. Express the subset as a filter over provenance fields (e.g. toolset == "api" and domain == "simple" and outcome == "pass"), evaluated over an allow-listed field set only (no arbitrary code). Resolve it to a named selection, compute the selection hash, and write a row manifest — never a loose id list pasted into the next command.
4. Contamination check + report. Before emitting, cross-check the selection against the frozen dev+holdout id sets. Detect: (a) holdout/dev ids inside a guarded selection; (b) the same task id appearing in both this selection and the frozen holdout; (c) duplicate trajectories (same content hash); (d) conflicting outcomes for the same (task_id, model, seed); (e) rows with missing provenance. Emit a contamination report; any guarded-pool violation hard-blocks unless overridden.
5. Emit the decontaminated downstream pool. On a clean (or explicitly overridden) check, write the final selection: a train-safe / distill-safe pool that provably excludes frozen dev+holdout, with its selection hash, the splits_sha256 it was tagged against, the corpus hash, the row count, and the row id manifest — so the claim survives audit. Hand this to the consumer.

Output Standard

End with:

• index path + corpus hash + per-source/per-split row counts;
• the named selection, its filter expr, its selection hash, and row count;
• the contamination report verdict: clean, blocked, or override (and if override, the logged who/when/why and exactly which guarded rows were admitted);
• duplicates, outcome conflicts, and missing-provenance counts found;
• the emitted decontaminated pool path + its hash + the splits_sha256 it cites;
• result_type: curated-selection or blocked;
• one recommended next consumer — compare-trajectories, local-distillation-lab (including its pedagogical arm), prepare-verifier-handoff (its stage-1 env authoring consumes the train-safe pool), or recursive-language-model pedagogical training — with the exact selection hash to pass it.

References

• ../capture-evidence/SKILL.md — freezes splits.json (train/dev/holdout); the split-membership source of truth this skill tags against and never re-derives.
• ../design-simulated-environment/SKILL.md — produces the trajectories/traces this skill ingests and curates.
• ../compare-trajectories/SKILL.md — consumes a clean selection to diff trajectories across models.
• ../prepare-verifier-handoff/references/stage-1-author-env.md — consumes the train-safe pool as the RL training set.
• ../prepare-verifier-handoff/SKILL.md — carries the selection hash + splits_sha256 into the hosted-RL handoff.