docs/curation/README.md

Curation Workflow

The curation step turns the pipeline-held-out curation_pool.jsonl into the two evaluation files shipped in data/splits/:

• test.jsonl — evaluation set. The paper's headline metrics come from this set.
• sanity.jsonl — 20-example smoke-test set for development.

In TRACE v1 the whole curation pool is promoted as the test corpus (no hand curation). The review + compile scripts are still part of the pipeline because (a) regenerating the test and sanity splits is deterministic from the pool, and (b) anyone adapting TRACE to a new clinical domain can reuse the same workflow with their own taxonomy.

---

Two scripts

Script	What it does
src/prepare_curation.py	Renders curation_pool.jsonl as a browseable Markdown document in docs/curation/review.md — one candidate per section, grouped by task × category, with gold labels and provenance visible.
src/compile_curation.py	Splits the curation pool into data/splits/test.jsonl (the remainder) and data/splits/sanity.jsonl (20 examples, largest-remainder stratified by category). Deterministic under --seed.

---

Commands

# Regenerate the browseable review document.
uv run python src/prepare_curation.py

# Compile the test + sanity splits from the pool.
uv run python src/compile_curation.py

# Override the sanity size (default 20) or seed (default 42).
uv run python src/compile_curation.py --sanity-n 20 --seed 42

After compile_curation.py runs you'll have:

• data/splits/test.jsonl — the evaluation set
• data/splits/sanity.jsonl — the smoke-test set

---

Reading review.md

review.md renders each candidate with:

• A heading showing the task type, category, and example_id.
• Gold labels inline (method, domain, level, learner profile, mastery state; or pattern class, behavior functions, escalation, confidence, crisis-plan flag).
• A short provenance line (for session interpretation: number of sessions, number of behaviors, whether IOA is included, whether ABC is included).
• The full user message (the teaching-program prompt or the session log).
• The full assistant message (the structured response).

Use LEGEND.md as a side-pane reference for the session-log notation in Task 2 examples.

---

Flagging issues for re-generation

If you spot a clinical inaccuracy while browsing review.md:

1. Note the candidate's example_id (shown in its heading).
2. Look up the candidate's meta.provenance.taxonomy_cells in the JSONL to identify which taxonomy dimension produced the issue.
3. Edit the relevant YAML under configs/ (or the renderer in src/generators/ if it's a rendering-logic issue) — never hand-edit individual JSONL entries, because the fix should propagate to every example that sampled the same cells.
4. Regenerate the corpus with src/generate.py --all, re-split with src/split_data.py, and re-compile with src/compile_curation.py.

Every clinical-review flag during v1 development landed as a single targeted taxonomy edit plus a full regeneration — never a hand-edit of the JSONL. This is the invariant the pipeline relies on: fixes must propagate through the taxonomy to be systematic across the corpus.

---

Adapting the workflow to a new clinical domain

The scripts are domain-agnostic. To port to a different dataset:

• prepare_curation.py just reads data/splits/curation_pool.jsonl and renders it. Works on any JSONL with the same envelope shape.
• compile_curation.py stratifies on category_of(example) — a small function near the top of the script. Change its logic to use whichever field best represents your category (the method field for TRACE's teaching programs; the pattern_class field for session interpretation).
• Everything else — largest-remainder allocation, deterministic splitting, provenance preservation — carries over unchanged.