A virtual molecule can fit on a screen and still be too large for the machine that claims to describe it. In quantum computing, the colourful picture of orbitals often arrives long before evidence of usefulness. Between a chemical geometry and an energy estimate lie decisions about electrons, basis functions, symmetries, circuits, measurements, and statistical correction. Each can introduce a legitimate failure. The ethical problem begins when a team learns that admitting it may damage funding, publication, or a commercial promise. An outlier is then removed without explanation, an inconvenient classical calculation vanishes from a chart, and an exploratory test is presented as a discovery. Trust does not come from declaring a fragile technology infallible. It starts when saying that a run did not converge, noise overwhelmed the signal, or the classical method won is not frightening. A virtual molecule earns public value only when a verifiable chain connects the chemical question, the computation, and a decision in the real world.
What a molecule represents inside a quantum calculation#
A real molecule is a system of interacting nuclei and electrons. A calculation does not insert all of it into a processor. It builds a model: it fixes or approximates nuclear motion, chooses a geometry, selects basis functions, and decides which orbitals and electrons belong in an active space. The electronic Hamiltonian is then translated into operators that an algorithm can estimate. Reducing the model is unavoidable; concealing the reduction is not. Two results carrying the same molecular name can answer very different questions when they use different geometries, charges, spin states, or basis sets.
A real molecule is a system of interacting nuclei and electrons. A calculation does not insert all of it into a processor.
In a variational method, a circuit prepares a parameterised state and the processor measures many terms to estimate an energy. A classical optimiser changes parameters and repeats the cycle. The final number contains model error, algorithmic error, sampling uncertainty, and device noise. Printing many decimals does not remove those layers. Nor is it enough to say that energy fell: the report must state the reference, number of measurements, stopping rule, and whether independent runs reached the same place. Useful chemistry also needs properties, forces, reaction barriers, or trends across candidates. One energy for a tiny instance can be a sound learning exercise, but it does not yet show that a material works.
A failure left unspoken becomes technical debt#
Informative failures take many forms. The optimiser may become trapped; the circuit may be too deep; calibration may change between batches; mitigation may reduce bias while increasing variance; a mapping may break a symmetry; or a classical comparison may show that the supposedly hard problem was easy at that scale. None of these observations proves by itself that the research lacks value. Each shows that the story needs boundaries. Recording the failure identifies what must improve and prevents another group from blindly repeating the same route.
Silence creates debt. The next person inherits parameters chosen without justification, charts without seeds or error bars, and a dataset in which only favourable runs survive. The team spends more effort sustaining an appearance than understanding the system. Psychological safety does not mean accepting careless work. It means separating professional dignity from the result of a test while holding the procedure to a demanding standard. Responsible leadership rewards early warnings, asks for an exact account of the harm, and prohibits retaliation for reporting uncertainty in good faith. Criticism addresses method and evidence; it never tries to silence the person carrying bad news.
A protocol before touching the processor#
The strongest defence against an opportunistic narrative is to agree on the protocol before seeing the outcome. Its record should define molecule, geometry, charge, multiplicity, basis, active space, and observable. It should also fix the classical reference, decision-relevant tolerance, circuit and measurement budget, and exclusion rules. If a legitimate reason for change later appears, the modification is dated and both analyses are shown. That distinction separates learning from retrospective selection.
The protocol includes positive and negative controls. A small case with a known solution checks the pipeline; a deliberate perturbation shows whether the metric detects degradation. Multiple seeds, measurement orders, and calibrations help estimate stability. The clock must cover preparation, queues, execution, mitigation, analysis, and verification, not merely the seconds when a chip is active. Cost includes access, supporting energy, staff, and opportunity. Finally, a provenance record links code, compiler version, topology, calibration, parameters, raw counts, and transformations. Without that chain, a molecular image may persuade, but it cannot be audited.
Case 1: the catalyst candidate that vanishes when the model expands#
A university laboratory studies a small molecule as a fragment of a catalyst. In a minimal active space, the quantum algorithm reproduces a reference energy, and the team drafts a release describing a promising candidate. A researcher checks a slightly larger space and finds that the order of two configurations reverses. The run consumes more resources, uncertainty grows, and the narrative advantage disappears. If reporting this means losing authorship or a contract, the group has an incentive to label the second test irrelevant.
Management establishes a review without punishment. It publishes the agreement in the small model and the reversal in the expanded one, identifies which missing orbitals explain the difference, and withdraws every claim about catalytic performance. The outcome becomes a validity map: the prototype tests an implementation at controlled scale, but it does not yet choose among chemical candidates. A higher-level classical comparison is added, and experimental verification is reserved for the point when several approximations agree within the preset threshold. The failure does not destroy the project; it prevents a limited numerical success from becoming a false recommendation about matter.
Case 2: mitigation that improves the average and hides dispersion#
A company tests a reference molecule over several days of remote access. It applies zero-noise extrapolation and obtains an average closer to the expected value. Some corrected runs, however, lie outside the physical range, and day-to-day dispersion doubles. The sales presentation retains only the final average and says mitigation restores chemical accuracy. An engineer notes that the procedure reduces bias under particular assumptions, but it does not make the hardware fault tolerant or guarantee stability.
The company rebuilds the report with paired raw and mitigated results, intervals, the rate of invalid runs, and sensitivity to extrapolation factors. It separates the set used for tuning from a held-out validation set. The conclusion changes: the technique helps in one configuration, while its variance still prevents its use for ranking molecules separated by small differences. The person who raised the warning joins the review and loses no responsibility. The asset gained is not a marketing sentence; it is a reproducible operating limit and a clear trigger for reassessment when calibration and usable depth improve.
Case 3: from a digital twin to the wet laboratory#
A public-private consortium wants to prioritise molecules for a capture material. Its workflow combines classical screening, a quantum subcalculation, and stability predictions. The ranking places a new molecule first. Instead of calling it a discovery, the consortium also selects a known control and several blinded candidates. An independent laboratory synthesises the compounds without knowing the predicted order. The leading molecule is difficult to manufacture and performs poorly in reality; another candidate from lower in the list works better.
The review finds three causes: the model omitted solvent, the relevant geometry changed during the process, and the calculated observable did not capture degradation. The consortium retains all data, updates its target, and publishes the discrepancy between ranking and experiment. It does not blame the quantum processor alone, because the whole chain contained classical and chemical decisions. Nor does it soften the result as merely awaiting validation. The real-world test reveals that the computational question was incomplete. The next cycle incorporates approximate free energy, synthetic feasibility, and stability, with separate thresholds. The case shows that a virtual molecule is not a product; it is a hypothesis that must survive independent measurements.
Separating four classes of error to learn properly#
Representation, algorithm, device, and inference errors should be distinguished. The first arises when a molecular model omits relevant degrees of freedom. The second comes from the ansatz, optimiser, truncation, or measurement strategy. The third includes noise, drift, readout, and imperfect gates. The fourth appears when a computed number becomes a chemical or commercial claim broader than the evidence. Mixing the four lets a team blame every discrepancy on hardware or, conversely, hide hardware limitations behind chemical complexity.
A useful report assigns tests to each layer. Convergence with respect to basis set and active space examines representation; ideal simulations and classical references examine the algorithm; interleaved repetitions and calibration records examine the device; external validation and decision analysis examine inference. Uncertainties do not always add simply, so assumptions and correlations must be preserved. The goal is not to manufacture one reassuring margin but to determine which conclusion survives. When the cause cannot be isolated, the honest result is indeterminate, accompanied by the smallest experiment capable of distinguishing the alternatives.
Governing incentives, data, and the right to stop#
A culture of trust needs mechanisms. Every project should have an incident channel, a reviewer independent of the commercial promise, and meetings in which anomalies are examined before successes. Employment metrics should not depend solely on clearing a technical threshold; they should also value reproducibility, documentation, and early detection. A decision log records who authorised changes and why. Where partners or funders are involved, the contract must protect publication of limitations and define how inaccurate public communications will be corrected.
The right to stop a demonstration is as important as access to the processor. It should be triggered when raw data are missing, a critical calibration changes, a control fails, a vulnerability appears, or uncertainty exceeds the relevant margin. Stopping does not mean cancelling forever: it preserves evidence, prevents premature decisions, and opens a time-bound review. Datasets need versions, permissions, retention rules, and traceable exclusions. Intellectual property may justify limiting chemical detail without erasing counts, method, or conflicts of interest. Legitimate confidentiality is not permission to invent certainty.
Four institutional sources that put boundaries around the promise#
The NIST places computing, networks, sensing, and metrology within quantum information science, reminding readers that measurement and standards are part of progress. The United States Department of Energy explains the principles and difficulties of quantum computing without confusing scientific possibility with a finished product. The United Kingdom's National Quantum Computing Centre provides resources on capabilities, adoption, and responsible development. The CERN Quantum Technology Initiative presents a programme combining research, collaboration, and training. These references help build questions and vocabulary; none certifies the performance of a particular molecule, algorithm, or vendor.
Institutional sources do not replace peer review, project data, or chemical expertise. They do help expose statements inconsistent with the state of the field: a simulation does not automatically inspect every configuration, a theoretical advantage does not guarantee end-to-end savings, and mitigation is not error correction. A source should appear beside the claim it supports, with an access date when the content may change. Procurement or policy also requires contractual specifications, independent testing, and comparison with the strongest accessible classical alternative.
A responsible decision keeps the negative result#
Before proceeding, an institution can ask: which real decision would change because of this number; what approximations are inside the virtual molecule; what is the baseline; what margin matters chemically; are all runs and exclusions reported; who can identify a failure without losing status; and what external test connects computation to matter? Answers should exist before scaling. If the quantum method ties or loses, the evidence set remains useful for teaching, improving tools, and preventing premature purchases.
The responsible conclusion may be to proceed, redesign, wait, or close. Each option needs criteria. Proceeding requires replication and an advantage tied to the task; redesign requires a falsifiable hypothesis; waiting requires a hardware or algorithm threshold that triggers review; closure requires archiving code, data, and reasons. Learning then survives the budget. Trust becomes a property of the process: the team can say what it knows, what it does not know, and which observation would change its view. The virtual molecule will have demonstrated value not when it shines in a slide, but when its prediction survives references, uncertainties, and matter. If it does not survive, reporting the failure precisely is a contribution, not a confession.
