Turns a candidate into a score you can trust. A reward number is only as honest\nas the variance around it and as the denominator under it: agents are stochastic,\nand infrastructure fails. evaluate produces a point estimate, its uncertainty, and\nthe count of tasks that actually produced a measurement. The math lives in\ncap_evolve.stats; this skill drives the adapter and aggregates.
A SplitResult (core/cap_evolve/loop.py:59-115):
reward — the mean, over the tasks that were scored, of each task's mean\nover its valid trials (harness.py:405-414, loop.py:127-131). Not the mean\nover every task in the split — see the next section.n_tasks / n_scored (and coverage = n_scored/n_tasks, loop.py:79-84) —\nthe honest denominator. Read these on every result, never reward alone.stderr — the combined SE of that reported mean: between-task variance\n(do different tasks agree?) folded with within-task trial variance (is the agent\nconsistent on a fixed task?), stats.combined_stderr. This is what the report\nprints and what the gate's significant mode consumes — not what the default\ngate reads; see \"What the gate actually consumes\".pass_k — when trials > 1, the estimated probability that all k i.i.d.\ntrials pass (reliability). Also pass_at_k — at least one of k passes\n(capability). Opposite questions; see references/concepts.md.diagnose reads:\n<run-dir>/rollouts/<split>/<task>__<tag>__t<k>.json (harness.py:334).The single largest honesty mechanism in the eval path. Two ways a trial produces\nno measurement:
\nrollout.error set) — the target never ran;rollout.error, so adapters must flag\nit by setting Score.raw[\"errored\"] (harness.py:308-323). An adapter that\ndoesn't is how a scorer outage becomes a real 0.0.Such a trial is excluded from the mean (harness.py:324-331), and a task with zero\nvalid trials is dropped from every statistic (loop.py:118-127). Averaging its\n0.0 in would state that the capability failed a task it was never given — which is\nhow a registry rate-limit storm produced val 0.000 and taught the optimizer to\n\"fix\" content that was never at fault. The rollout file is still written, for\nforensics.
What to check. raw.valid_trials == 0 on a per-task record means unmeasured,\nnot failed. A reward computed over a third of a split describes the\ninfrastructure, not the edit. Below coverage 0.6 the gate returns\nindecisive=True and declines to judge rather than calling it a regression\n(gate.py:137-146) — a run producing repeated indecisive steps has an\ninfrastructure fault, not a bad optimizer. Pinned by\ncore/tests/test_infra_errors_not_zeros.py (518 lines).
When per-task data is available the loop sets gate mode to paired\n(harness.py:1524-1526), and paired mode recomputes the SE from the per-task\ndeltas against the same tasks (gate.py:156-160); SplitResult.stderr is never\nread on that path. So what extra trials buy you under the default gate is a more\nstable per-task mean, which shrinks the paired delta variance — not a smaller\nstderr. stderr feeds the report and the significant fallback used when\npaired data is unavailable (gate.py:184-207).
python scripts/run.py --run-dir .capevolve/run_XXXX --project .capevolve/project \\\n --candidate seed --split val --n-trials 3\n--split accepts only train or val, enforced by argparse choices\n(scripts/run.py:25) — a --split test invocation exits non-zero. The\nenforcement lives in this CLI, not in harness.evaluate_candidate (issue #361),\nso never \"helpfully\" widen those choices.--n-trials defaults to 1. On a stochastic target that is the degenerate\ncase below; run.py prints a warning to stderr when it happens.--ks picks the k values for passk; it defaults to 1..n_trials, so\n--n-trials 3 reports pass1..pass^3. Any k above a task's trial count is\nomitted rather than reported as 0.0 (loop.py:134-147).CAPEVOLVE_WORKERS=N generates rollouts through a thread pool\n(harness.py:49-57); scoring stays serial so the numbers match a serial run.\nKeep it at 1 if run_target is not thread-safe (shared scratch dir, one live\ncontainer, module-global client) — harness.py:225-227.ids=) is never gateable: its n_tasks is the subset,\nso coverage reads 1.0 (harness.py:229-237).Two axes, and the trials axis is the one people get wrong.
\nk runs with\nseed = base_seed + k (harness.py:374, trials.py:10-13) and the adapter\ncontract requires passing it to a stochastic runner (adapter.py:52-54). An\nadapter that drops it gives you n identical copies: per-task stderr is 0,\npass^k is exactly 0 or 1, and the whole apparatus looks healthy while measuring\nnothing. cap-evolve check can prove it: with\nCAPEVOLVE_N_TRIALS=3 CAPEVOLVE_CHECK_TRIAL_PROBE=1 it fires two real rollouts at\ndifferent seeds and warns if they are byte-identical\n(core/cap_evolve/check.py:169-190). It is opt-in because the probe costs real\nrollouts — run it once per adapter, and treat the warning as \"every variance number\nhere is fiction\". Trials cost budget linearly, so spend them where variance actually\nthreatens a decision — the val split the gate reads — not on every exploratory probe.stats.stderr returns 0.0 below 2 tasks and combined_stderr's\nbetween-task term is 0 below 2 (stats.py:28-30, 47-50), so a 1-task val gives\nstderr = 0, a bar of 0, and the gate degenerates to strict (\"any Δ>0 wins\") with\na logged warning (gate.py:40-60). Below roughly 5 val tasks the k·SE bar is\ndominated by sample size and is optimistic — issue #113. An empty val presents as\ncoverage 1.0 with reward 0.0 (loop.py:79-84).A one-task gain is not reliably bankable. Under the shipped default\n(mode: paired, k_se: 1.0) a candidate that improves exactly one val task and\nchanges nothing else has Δ̄ == SE(Δ) algebraically, so the strict > at\ngate.py:176 is settled by floating-point representation — rejected at n=4, 8, 50,\naccepted at n=20, identical printed numbers. Issue #351, open; derivation in\nreferences/concepts.md. Do not read a rejection of a single-task fix as evidence\nthe edit was bad — check how many tasks moved.
n_trials ≥ 3 on a stochastic agent with the seed forwarded; stderr\nnon-zero; n_scored == n_tasks; pass^k inspected alongside the mean.coverage first, always); single-trial scores feeding a\nsignificance gate; identical trial rewards across seeds; trusting a high mean when\npass^k is low (the gain is fragile).references/concepts.md (125 lines) — the variance decomposition and the\ncombined-SE formula, where these statistics break down on small samples (including\nthe #351 Δ̄ == SE derivation), pass^k vs pass@k with their unbiased estimators,\nbootstrap CIs, where the test-split refusal is enforced, and sources. Load it when\nyou need the statistics themselves rather than how to run an evaluation.