Distill Test Set

Turn a large, noisy dataset into a small, high-signal Coval test set. Big datasets are the wrong input for simulation: running thousands of sims is slow, expensive, and likely to overwhelm the agent under test, and most rows are duplicates, paraphrases, or boilerplate. The goal is a representative 10–50 case suite that covers the agent's real scenarios once each, weighted toward known failures.

Drive everything through the coval CLI and the bundled scripts/distill.py. Follow the phases in order. Present a summary and get confirmation before creating anything.

If $ARGUMENTS is a Coval test set id (8 chars) or a file path, use it as the source in Phase 1.