Scientific CI/CD for Self-Modifying Discovery Agents: Statistical Gödel Gates, Capacity Budgets, and Domain Verifiers

Self-evolving scientific agents can improve by rewriting prompts, tools, workflows, memory, search policies, and eventually evaluators or learned parameters. This capability creates a deployment bottleneck that ordinary benchmark reporting does not address: an edit can improve visible scores while silently degrading calibration, robustness, scientific validity, or future learnability. We propose Scientific CI/CD, a governance architecture that treats every self-modification as a high-risk pull request. A candidate edit is promoted only after passing three gates: (i) a Statistical Gödel Gate using anytime-valid e-values or confidence sequences on protected holdouts; (ii) a Capacity Budget Gate that prices model, tool, memory, data-access, and autonomy expansion; and (iii) a Domain Verifier that checks conformal calibration, scientific invariants, and claim-evidence provenance. We define accepted-edit regret, lifetime harmful-adoption rate, evaluator-corruption gap, calibration preservation, rollback utility, and cost-adjusted verified discovery. Three experiments instantiate the gates: a sequential self-edit stream, a capacity-expanding model-selection benchmark, and a dry-lab RNA-seq twin. Across 200-proposal streams, naive repeated testing accepted harmful edits in 6.1% of promoted changes, while the spending-based e-Gödel gate reduced this to 0.008%. Capacity budgeting reduced hidden-risk overfitting in self-expanding polynomial agents, and the domain verifier rejected harmful biomedical workflow edits that passed mechanical or statistical checks. The result is a concrete merge contract for self-evolving scientific agents: improve, but only under valid evidence, bounded capacity, domain constraints, reproducible provenance, and rollback.