gate — accept only real improvements, on val

The gate is where dishonest optimization is prevented. Search is a noise amplifier: try enough candidates and some will look better by chance alone (the more candidates you screen, the larger the expected best-of-noise). The gate is the rule that keeps a lucky draw from being promoted to "the new best". It refuses any split but val, and by default accepts a candidate only when its val reward beats the current best by more than k standard errors.

Inputs / outputs (manifest tokens)

• needs: scores — the candidate's and current best's val reward and stderr (from evaluate). The SE is not optional: significance is meaningless without it.
• provides: decision — {accept, reason, delta, threshold}, the audit record of why a candidate was kept or rejected.

The significance rule

paired (the default):   accept ⟺ mean(Δ[t]) > k · SE(Δ)     over the SAME val tasks
significant (fallback): accept ⟺ Δ = cand − curr > k · sqrt(cand_se² + curr_se²)

The bar is Δ > k·SE and not Δ > 0 because search is a noise amplifier: screen enough candidates and the best-looking one is best by luck, so Δ > 0 banks noise as progress and the val curve climbs while nothing improved. Clearing k standard errors of the measurement's own error is what makes an accept mean something — turn this down and the run's numbers stop being evidence. k=1 is lenient (~1σ); raise it to be stricter. It is the textual-optimization analogue of Koehn's bootstrap significance test for metric differences.

paired is stronger because both sides were scored on the same val tasks, so per-task difficulty cancels and only the paired variance counts; significant treats the two means as independent samples and is only correct when they are.

Single-trial scores report stderr=0, collapsing k·SE to 0 — then significant silently degrades to strict and accepts any positive blip. If you run the significance gate, score with multiple trials (see evaluate).

Modes

• paired (the default): mean(per-task Δ) > k·SE(Δ). The loop selects it whenever per-task val data exists (harness.py:1524-1526, gepa.py:741-743) and capevolve.yaml ships gate_mode: paired.
• significant: Δ > k·SE_combined — the unpaired fallback, used when the two sides aren't aligned per task. decide()'s own mode= parameter defaults here for bare callers with no per-task data; that is not the default of a real run.
• threshold: Δ > T — a flat margin (use when you have a domain minimum worthwhile gain, e.g. "don't bother unless +2pp").
• strict: Δ > 0 — any improvement. Only safe with a near-zero-variance scorer (deterministic, single correct answer).

Anything else raises. There is no simplicity/size mode: it was unreachable dead code (nothing ever supplied a size) so it silently behaved as strict, and it has been removed rather than documented.

No-regression (the second gate)

A mean can rise while previously-passing tasks silently break. Pair the significance gate with a no-regression check: reject a candidate that improves the aggregate but drops any task that the current best passed. This is the same dual-gate discipline SWE-bench-style harnesses use (a patch must pass the new tests and not break the existing ones — FAIL_TO_PASS and PASS_TO_PASS). diagnose provides kept_good (the currently-passing tasks) precisely so this check has something to protect.

Dual-mode

This phase runs two ways from the same SKILL.md: standalone as the slash command /cap-evolve:gate (the argument-hint shows its run.py args), and orchestrator-callable — cap-evolve run / the orchestrate skill invokes the same scripts/run.py headlessly and threads the run dir between phases.

How to run

python scripts/run.py --current 0.50 --candidate 0.62 \
    --mode significant --k-se 1.0 --candidate-stderr 0.03 --current-stderr 0.03

Algorithms call the gate internally every iteration via the harness; this skill exists so a human or agent can reproduce and understand a single decision.

What good vs bad looks like

• Good: paired mode (the default) with real multi-trial SEs; a no-regression check on top; every accept/reject logged with its reason.
• Bad: gating on train (the tool refuses this — it overfits the optimizer to the data it edits against); strict mode on a noisy agent (accepts noise); raising the mean while quietly regressing tasks because no-regression was off.

References

• references/concepts.md — the difference-of-means SE, choosing k, the multiple-comparisons motivation, the dual-gate / no-regression rationale, and why gating on val (never train, never test) is the honest split, with sources.