evolve: skip no-op generations, guard re-scores, cap run budgets

[Evanfeenstra]

What went wrong

Reviewing gaia-evolve run 1788061734710 (prod, still running at 11h): generation 5's author degenerated into a filler loop — 56KB of </br>proceed.</br>meta_run_step</br> and zero-width spaces — and published nothing, echoing garbage into its structured output:

version: "call memeta pickameter="
summary: "}false"
changes: []

The vpin.version || vactive.version fallback did its job and refused the garbage echo. But it fell through to the candidate's active version — which was generation 4's own publish, v11. So generation 5 spent ~2.5h and ~$8 re-running an already-graded version over all 25 tasks.

Worse: exact-match fitness is resampled. That re-run of v11 was 13/13 at the time of review, against v11's recorded 0.76. evolve-loop has no duplicate-version check, so it was on track to write gen 5, v11, improved into the lineage — produce-sampling noise recorded as a hill-climb step, and a final report naming a best version at an inflated accuracy.

#1620 hardened the summary channel of this same author failure. This is the version/grading side of it, plus the budget problem the run exposed.

Fixes

1. Skip, don't re-score — gaia-evolve-gen.yaml, harvey-evolve-gen.yaml

A new vbefore step reads the candidate's active version before the author runs (with an ordering edge into author so it can't race the publish). A published if-gate compares it to the resolved version afterward; when they match, nothing was published, and candeval is skipped entirely. result fans in on the gate as well as the digest — vein only skip-propagates when every dep was skipped — so it still reports, carrying noop: true.

A no-op generation now costs one author instead of a full task set. At generation 0 the candidate doesn't exist yet, so vbefore.version is undefined and any real publish trips the gate true.

2. Guard the best-update — eval/evolve-loop.ts

• No-op generations record no fitness. Writing a 0 would libel an approach that was never tried and push later generations to explore away from it; the briefing now says NO CANDIDATE PUBLISHED instead.
• isNewBest + a scored ledger: a version this run already graded cannot become the best on a second, luckier sample. The gate above stops the common cause, but a deliberate republish of identical YAML under a new version string is indistinguishable from the loop — this is the backstop, and it pins the reported best to the run where that version was first measured. Applied on both the live and journal-replay paths. The resample is still reported honestly, flagged rescoredVersion.

3. Budget caps — maxCost / maxMinutes

Checked between generations, so they bound when the loop stops and never kill a generation mid-flight. Both null (uncapped) by default, wired through gaia-evolve / harvey-evolve params and overridable per run.

Generation count alone is a poor budget: authors reliably evolve toward more expensive architectures (dual independent attempts + a reconciler, by gen 3). The observed run drifted from ~1h/gen to ~2.7h/gen and $16 to $32 cumulative — a "10 generation" run headed for ~22h and ~$65.

Testing

tsc clean. Both evolve smokes pass, with new cases for:

• the published gate expression across all four version states, including the exact live failure (vbefore v11 / bogus pin / vactive v11 → no-op) and undefined-safety against meta/get-workflow's { error } return
• the no-op branch: no fitness recorded, best untouched, only the author budget spent, briefing wording
• the re-score guard: same version at 0.6 then 0.9 keeps bestFitness at 0.6
• maxCost stopping between generations, and absent caps changing nothing (incl. null params parsing as uncapped)

Note on the live run

The in-flight run is on the old code and safe to let finish — v10/v11 are real gains over the 0.68 baseline. But distrust any "best" it reports at v11 from generation 5 onward; that's the resample, not a climb. Held-out validation before promoting was already the standing rule here and still applies.