Simulated benchmark dataset: hidden-cause discrimination in a passive air-breathing PEM fuel cell from a declared-amended dynamic model (700 runs, v1.1 sealed)

All signals in this record are simulated; it contains no experimental measurements. The generator is a declared-amended version of the dynamic air-breathing PEFC model of Calili et al. (2021), Int. J. Hydrogen Energy 46, 17343–17357: the oxygen state, thermal state, and oxygen-sensitive activation kinetics are amendments introduced for benchmark construction and are not attributed to the original model. The Simulated Passive-PEFC Hidden-Cause Discriminator Benchmark provides exact hidden-state ground truth for developing and testing hidden-cause observers in passive air-breathing polymer electrolyte fuel cells. It contains 700 simulated runs across 140 trajectory cells (five load profiles × five noise seeds per condition), spanning seven branches: a nominal control, pure oxygen-transport stress, pure dry-out/ohmic stress, thermal and anode-side branches, and mixed, guard, and confound cases. Two data layers are provided: Layer 1 (full model variables — exact hidden states) and Layer 2 (sensor-facing signals with a declared noise model and nonuniform sampling). Ground truth is exported as multi-hot cause labels and continuous hidden-state labels. Data splits are fixed at trajectory-cell level to prevent near-duplicate leakage; the central intended task is pure-to-mixed compositional generalization, with oxygen-component recovery inside mixed runs as the principal open problem. Files in this record: pefc_benchmark_v1_1.zip (ACTIVE sealed release; SHA-256 40ae52eaa51bbeceb8c324ed873f89c263de5fe47d81552bfe9a7dfdd12d4c05), pefc_benchmark_v1_0.zip (ARCHIVED prior release, retained for lineage; SHA-256 296768b02be3a5636332b21d0f52545d5c287dfab1caac6e04aad337cdab221a), baseline_package_v1_1.json (version-aligned baseline results, thresholds tuned on the v1.1 validation split only), fig8_baselines_v1_1.png, and DEPOSITION_README_zenodo.md. Both sealed archives are self-verifying: every member file's SHA-256 prefix is recorded in the in-archive generation log. The associated data-descriptor manuscript is intentionally not included in this record; it will follow separately once authorship is finalized (its seal-time fingerprint is retained in the enclosed README for lineage). Corrective findings are part of the record rather than hidden: the first observer designed against the benchmark produced a decisive negative result, and the v1.0 validation split was found structurally unsuitable for fast-edge models and repaired in the explicit v1.1 release; v1.0 is retained because the manuscript's matched-amplitude audit and first-contact results were computed against it and are labeled as such. Non-license: this dataset makes no claim of sim-to-real transfer, experimental validation, flooding physics, validated thermal behaviour, novelty, or priority. The binding non-license text is inside the v1.1 archive (README_non_license.md). The temperature treatment shares a provisional element with the companion record: the published Appendix B coefficients of the source paper do not reproduce its figures. This dataset is the simulation arm of a two-arm research programme; the experimental arm is the pre-registered matched-endpoint future-probe design deposited at 10.5281/zenodo.21497865. If hidden predictive state is confirmed experimentally, this benchmark is the observer-development environment for it; if not, it stands alone as a methods object for hidden-cause observer research. Developed via a human-mediated cross-model workflow ("Research Compass"): sequential adversarial exchange between two AI systems (OpenAI ChatGPT and Anthropic Claude), orchestrated, audited, and frozen by the author. A methods report documenting the process, including errors caught and corrected, is in preparation.