Evidence-Gated Dynamic Campaign Control for Materials Discovery

Self-driving laboratory software usually asks an optimizer for the next candidate while leaving the objective, search space, experimental route, decision policy, and campaign mode fixed, even when execution invalidates that specification. We present HELIOS, a typed control plane that separates proposal, validation, authority, execution, and replay during evidence-gated campaign revision. Role-scoped services need not be language-model agents; an optional model can only supply a typed proposal. We release an anonymized MIT-licensed reference implementation containing TaskContract schemas, a capability manifest, evidence and decision ledgers, replay utilities, and a fully reproducible 200-cell matched benchmark. In that study, execution-aware route re-ranking reduced modeled switching cost from 11.6 to 7.1 units (paired difference 4.5, 95\% CI 2.7--6.4; Holm-adjusted $p=0.0014$). Endpoint benefit was unresolved and regret-AUC was descriptively worse, exposing a cost--search trade-off rather than universal improvement. Retrospective HER, liquid-handling, and solubilisation vignettes illustrate policy, execute--observe, and fail-closed interfaces; they are reported separately from the matched finding.