Datasets:
Preregisterable mechanism and decision validation
Scope
Single-qubit pure dephasing with independently calibrated binary integrated phase magnitudes. No claim that arbitrary natural noise has this support. Every prefix must use the marginal of a common environmental preparation law.
Protocol A: expose the hidden direction
Freeze the seven-bin fourth-order sign laws p_plus and p_minus, common amplitudes, all prefix controls with at most two pulses, and the three-pulse target cuts (1,3,5). Calibrate phase magnitude, pulse width/error, relaxation, state preparation, and readout independently. Generate the sign histories from the released exact law, not by fitting the acquired data. Interleave the laws and measurement settings where the preparation protocol permits it. Log chronology and all failures.
Primary predictions: low-pulse responses agree within the propagated error budget; the target contrast follows the exact response product. Scan a common amplitude as a control: the low-pulse difference stays zero in the ideal model. Add a local pair amplitude imbalance to test the explicit capability tradeoff. Distinguish phase error from a failure of the hypothesized common law.
Protocol B: certify before measuring the target
Use six bins at independently validated phase magnitude pi/4. Acquire prefixes (2,()), (4,(1,)), (6,(1,3)) and baseline (6,()). Freeze the simultaneous confidence rule and systematic allowances. Publish the target lower bound and paired gain before measuring target (6,(1,3,5)). The target data are validation, not selection input. Include the extra target pulse's independent error allowance.
Baselines and costs
Compare a direct target/baseline scan, the proposed prefix certificate, and an appropriate established control-adapted method under matched support priors, confidence, number of shots, readout, pulse imperfections, reset costs, and target fidelity. Record total duration and required host processing. Do not substitute settings for shots or ideal pulse counts for total time.
Failure conditions
Reject the device-model claim if the calibrated model cannot jointly explain prefix, baseline, and held-out target data. Reject a confidence claim if coverage violations exceed the preregistered combined statistical/systematic budget. Reject an efficiency claim if its gain disappears against a real matched competitor. Empty feasible sets produce an abstention and diagnosis, not a favorable fitted law. No sample size is universally prescribed; the supplied synthetic 50,000-shot example is not a laboratory power calculation.