HilbertBench Study E: Blinded, Pre-Registered Diagnosis of Planted Quantum Machine Learning Failure Modes — Corpus, Answer Key, and Scores

The complete data package for the first blinded, pre-registered validation of a quantum machine learning diagnostic. A corpus of 72 recorded QML runs with planted failure modes (healthy, barren plateau, shot starvation, noise domination; 36 simulated, 36 executed on IBM quantum hardware) was blinded and diagnosed by an independent researcher using only the HilbertBench instrument's outputs, then scored against a SHA-256-committed answer key in a recorded unsealing. Result: 72/72 correct (95% Wilson CI 94.9–100%, p ≈ 4×10⁻⁴⁴ vs 25% chance), with the pre-registered hardware barren-noise co-occurrence confirmed 9/9 in secondary labels. Contents: the blinded corpus (sealed, verifiable traces), the answer key and its pre-published commitment, the diagnostician's frozen sheet and its commitment, scoring output, ground-truth manifest, the unsealing session typescript, diagnostician instructions, and a README with a full verify-it-yourself recipe. Instrument: HilbertBench v1.0.0 (DOI 10.5281/zenodo.21009245). Pre-registration: OSF, DOI 10.17605/OSF.IO/2PWHU.