Reproducible Quantum Research
Reproducible quantum research makes it possible to reconstruct a calculation and test whether its scientific conclusion survives independent implementation and reasonable changes of assumptions. The frontier is not another archive of scripts, but reliable communication of models, uncertainty and dependencies across complex workflows. Bitwise reproduction and scientific agreement answer different questions.
Required background. Convergence reports and error budgets provide the numerical contract.
Preserve the claim, not only the executable
Section titled “Preserve the claim, not only the executable”Record the Hamiltonian or channel, units, basis ordering, boundary conditions, truncations, initial state and observables. Preserve raw outputs and the transformation that makes each figure. Pin the environment sufficiently to reconstruct the run, while specifying numerical acceptance criteria that tolerate legitimate hardware-dependent rounding.
For an observable , a meaningful reproduction criterion might be
provided the terms are defined consistently. Correlated statistical errors need covariance-aware treatment; unknown systematic bias cannot be turned into a small error bar by rerunning the same code.
Established practices and active research
Section titled “Established practices and active research”Wilson and colleagues describe practices that improve scientific software reliability. FAIR principles address findability, accessibility, interoperability and reuse. Established: these practices make evidence easier to inspect. Active: automated provenance, portable scientific tolerances, uncertainty propagation and independent validation for heterogeneous quantum workflows. FAIR metadata or a passing test suite does not prove the physical model. Software practices, FAIR principles.
Containers preserve an execution environment but can conceal undocumented inputs. Independent implementations reveal some shared assumptions but can repeat the same derivation error. Analytic limits, conservation laws and experimentally validated observables supply complementary checks.
Open questions
Section titled “Open questions”How can a workflow record the provenance of adaptive decisions and failed runs? Which tolerances remain meaningful across precision, hardware and solver changes? How should stochastic artifacts support repeated analysis without encouraging seed selection? What minimal package lets another group reconstruct the claimed uncertainty rather than only the central value?
Graduate project: reproduce one figure independently
Section titled “Graduate project: reproduce one figure independently”Choose a computational paper with public code and a modest system size. First reproduce a figure using the released environment. Then implement one essential observable independently and perform a convergence study. Deliver a compact reproduction package, a discrepancy log and a statement of which scientific claim is supported. A laptop or workstation and publicly accessible data are sufficient; avoid projects whose core inputs are unavailable.
The baseline is the authors’ documented run. Failure tests alter basis ordering, unit convention, random seed and solver tolerance. Keep tests that detect real scientific mistakes rather than tests that merely repeat the implementation. Research begins when a recurring failure mode motivates a reusable provenance representation, numerical acceptance method or statistically justified reproduction protocol.
Entry points and scope
Section titled “Entry points and scope”Read the software-practice paper, then inspect one real result’s dependencies. Sources checked on 11 October 2026; selected methodological anchors are 2014–2016. A new tool merits adoption when it reduces an identified reproducibility failure under independent use, not merely because it can package more files.