Negative Results Matter: Where Q-Psi Did Not Beat the Baseline
1. Publication Bias in Quantum Computing
In contemporary quantum software research, negative or inconclusive results are routinely concealed. When a new heuristic fails to beat an industrial baseline, researchers often tune hyperparameters post-hoc until a positive graph is found.
Q-Psi operates under a pre-registration protocol: hypotheses, compiler seeds, and verification metrics are committed to Git before QPU submission. When physical results falsify our initial hypothesis, the results are preserved in full.
2. Four Preserved Negative & Inconclusive Studies
We hypothesized that single-pass calibration graph weighting would outperform Qiskit stochastic layout search. Physical execution on IBM Marrakesh showed 1 win, 2 losses, and 3 inconclusive outcomes (mean delta \(-0.0106\)). Qiskit's multi-seed stochastic search remained superior or competitive.
Direct fractional pulse synthesis compressed schedule durations by up to 40%, but translated into physical fidelity wins on only 2 out of 6 circuits (4 cases were inconclusive).
Risk-mass layer selection successfully reduced twirled layers by 58.9%, but non-inferiority compared to full randomized compiling was not statistically established across 4 cases.
Secret periods were recovered on 6/12 instances up to \(N=28\) (56 physical slots), but two-qubit gate error accumulation prevented establishing asymptotic physical speedup.
3. The Scientific Value of Negative Baselines
Preserving negative comparators gives authentic meaning to positive results (such as our Grover candidate search and Dynamic BV advantages). A research program that only publishes positive graphs is indistinguishable from cherry-picking.