Dataset for Certifying adaptive experimentation from decision-time information

This dataset contains the frozen evidence, publication source data and reproducibility materials accompanying the manuscript “Certifying adaptive experimentation from decision-time information” by Jia Bi and Chenyang Zhu. The manuscript introduces OPAL, a decision-theoretic framework for determining whether adaptive experimental capacity is supported by the information available at decision time. The framework separates latent adaptive opportunity, deployable opportunity, finite-campaign certifiability, safe fallback and lifecycle adoption. It is designed to distinguish settings in which adaptive opportunity is absent from those in which opportunity exists but cannot be identified reliably using currently available observations. The deposit contains: publication-ready main and Extended Data figures; CSV source data underlying every reported figure; the main manuscript, Supplementary Information, bibliography and plotting programs; a frozen retrospective analysis of selector aggressiveness, execution value, information cost, probe acquisition and model-complexity fallback; exact finite-population certification calculations and the corresponding finite-campaign validation evidence; held-out simulator calibration and independent risk-control validation; descriptive constructibility analyses based on measured Causal Chambers trajectories; prospectively partitioned development, calibration and locked evaluation outputs derived from measured Cancer Therapeutics Response Portal (CTRP) responses; result-to-source mappings, evidence-status audits, governance records and SHA-256 manifests. The CTRP evaluation uses complete measured response coverage for a frozen four-compound panel. Values for the prespecified capacity actions are therefore obtained through measured-outcome lookup rather than propensity weighting, outcome imputation or simulated counterfactuals. The evaluation is an offline assay-capacity analysis and does not constitute an online treatment recommendation, clinical validation or prospective laboratory deployment. The Causal Chambers component establishes that the observable score, action and fallback computation can be instantiated reproducibly on archived physical-system trajectories. Because its action mapping is author-defined and the analysis was specified after archive inspection, it is provided as descriptive constructibility evidence rather than causal physical-policy validation. Third-party raw CTRP and Causal Chambers archives are not redistributed. The deposit instead provides source URLs, pinned versions, checksums, frozen partitions, derived ledgers and validation reports needed to trace the reported results back to their public sources. A separate linked Zenodo record, “OPAL: software and reproducible analysis code”, contains the source code, configurations, protocols, tests, validators and executable runners. Evidence roles and claim boundaries are preserved explicitly so that retrospective analyses, exact certificates, simulator validation, physical-data constructibility and locked measured-outcome evaluation are not assigned the same inferential status.