baseline is the first phase that touches data, so it owns the run's one\nirreversible decision: the split. It writes splits.json once (seeded), scores\nthe unmodified seed capability on val, and records that score as the bar every\nalgorithm must beat.
Run implement-and-check first. baseline re-runs that check itself and exits\nnon-zero before creating a run dir if it is red — a split frozen against a broken\nadapter poisons every number measured afterwards.
headroom (1 - val) and\nheadroom_verdict: saturated means the seed is already at the ceiling and\nfurther iterations buy noise — stop; floor (val at 0) usually means a\nmis-wired adapter rather than a hard task — re-check before spending budget;\nok means proceed. The same verdict is logged as a headroom event so the\norchestrator can stop on it with no human reading the number.0.5 / 0.25 / 0.25): deterministic given\n--seed. Reproducible runs partition identically.--split-ids): a JSON {train,val,test} of ids — use a\nbenchmark's official split, or set all three equal to fit the whole set with\nno holdout (the test number is then a fit metric, not a held-out result;\nthe run dir records a splits_warning saying so).n, and\na candidate that improves exactly one val task cannot reliably clear it at all\n(issue #351), so size val with the decisions it has to make in mind.--reuse-baseline PRIOR_RUN_DIR)Re-scoring the seed is wasteful when the split + seed are unchanged.\n--reuse-baseline <prior run_* dir> (spec key reuse_baseline) copies that run's\nsplits.json, baseline.json, seed snapshot and seed val rollouts into the fresh\nrun dir and skips the baseline eval; the copied test_used flag is reset so this\nrun can still finalize on test exactly once. --resume is the same-run variant:\nreopen an existing run dir, skip the eval when baseline.json is already there.\nBudget flags (--max-iterations, --stall, --max-usd, …) are accepted here\nbecause the run dir owns the budget and later phases read it from there.
Runs standalone (/cap-evolve:baseline) or headlessly via cap-evolve run; same\nscripts/run.py either way.
python scripts/run.py --base .capevolve --project .capevolve/project \\\n --capability seed_capability --seed 0 --ratios 0.5,0.25,0.25 \\\n --max-iterations 10 --stall 2\nPrints the run-dir path (used by the algorithm + finalize), the split sizes, the\nbaseline val and the headroom verdict. Use --n-trials ≥ 3 for stochastic\ntargets so the baseline carries a real stderr rather than 0.
The one failure mode nothing later can repair is re-splitting after this phase: a\ntask migrating out of test leaks the held-out set, and every later number —\nincluding the sealed test score — becomes unfalsifiable.
\nreferences/concepts.md — why the split is frozen once and seeded, how to read\nthe headroom verdict, and why no-holdout runs are fit metrics, with sources.