Quantum Proof-of-Concept Exit Criteria: A Scorecard for Results, Readiness, and Next Steps
A quantum proof of concept should end with a decision, not just a demonstration. This guide provides a repeatable scorecard for evaluating the problem, classical baseline, data, circuit, hardware, vendor, economics, reproducibility, and business relevance so an enterprise team can decide whether to stop, repeat, expand, or close a quantum experiment.
Overview
A quantum computing POC is an investigation under constrained conditions. Its purpose may be to test a business hypothesis, compare quantum hardware or software platforms, develop internal skills, or determine whether a particular workload deserves further investment. Those goals require different measures of success. A circuit that runs successfully on a simulator may demonstrate programming progress without proving business value. Likewise, a hardware experiment may produce useful engineering evidence even when it does not show quantum advantage.
The exit review should therefore separate four questions:
- Did the team answer the technical question? For example, can the workload be encoded, executed, measured, and reproduced?
- Did the experiment produce credible comparative evidence? The result should be assessed against an appropriate classical method, not an undefined notion of “faster.”
- Is the selected hardware or vendor suitable for the next stage? This includes access, workload fit, reliability, tooling, controls, and total effort.
- Is there a reason to continue? The next phase should have a defined hypothesis, budget, owner, and evidence threshold.
Use a weighted scorecard rather than a single pass or fail label. A low score can mean the POC should close, but it can also identify a specific weakness to correct in a second experiment. This is especially important when comparing platforms with different qubit technologies, execution models, connectivity, error characteristics, programming frameworks, or access arrangements. A larger nominal qubit count alone should not determine the decision.
For hardware-specific context, pair this framework with a practical quantum hardware benchmarking method and a quantum computing vendor RFP template.
How to estimate
Start by defining the decision the POC must support. Examples include selecting a platform for a follow-up experiment, deciding whether a workload is ready for hardware testing, determining whether a hybrid algorithm merits optimization, or closing an investigation because the expected value is too low.
Score each category from 0 to 5:
- 0: No usable evidence or the criterion was not addressed.
- 1: Initial work exists, but it is unreliable or poorly defined.
- 2: Partial evidence with important gaps.
- 3: Adequate evidence for the current POC stage.
- 4: Strong evidence that supports a next-stage decision.
- 5: Repeated, well-documented evidence suitable for a formal expansion decision.
Then apply weights that reflect the purpose of the project. A general enterprise scorecard can use the following starting point:
| Category | Suggested weight | Exit question |
|---|---|---|
| Problem definition | 10% | Is the business or technical question precise and measurable? |
| Classical baseline | 15% | Is there a credible method for comparison? |
| Data and workload preparation | 10% | Can representative inputs be prepared without hiding material costs? |
| Circuit and algorithm performance | 15% | Does the implementation produce the intended output within stated limits? |
| Hardware and vendor fit | 15% | Does the platform support the workload and the next experiment? |
| Reproducibility | 10% | Can another team rerun and interpret the experiment? |
| Economics | 10% | Are time, access, engineering, and operational assumptions visible? |
| Business relevance and risk | 15% | Does the evidence justify a defined next step? |
Calculate each weighted contribution as score divided by 5, multiplied by the category weight. Add the contributions to produce a result from 0 to 100. Treat the number as a decision aid, not a universal industry benchmark. Set local gates before reviewing the result. For example, a team might require a minimum overall score while also requiring at least 3 out of 5 for the classical baseline, hardware fit, and reproducibility. A high average should not conceal a critical failure in one of those areas.
Use four outcomes:
- Close: The hypothesis is not supported, or the remaining value does not justify more work.
- Repeat: The question remains valid, but a defined technical gap prevents a reliable conclusion.
- Expand: Evidence supports a larger workload, more representative data, another platform, or a deeper hardware comparison.
- Production-readiness review: The POC has evidence worth assessing against operational, security, governance, and support requirements. This is not the same as production approval.
Inputs and assumptions
Problem and baseline
Write the target outcome in operational terms. “Use quantum computing for optimization” is too broad. A stronger statement identifies the decision, constraints, input scale, output quality, and time horizon. Record what counts as a useful result before examining the quantum output.
The classical baseline must be appropriate to the problem and available at comparable input sizes. Record the algorithm, implementation, hardware, preprocessing, postprocessing, runtime boundary, solution-quality measure, and tuning effort. If the POC compares a quantum circuit with a simple classical script, label that limitation clearly. Do not call an improvement “quantum advantage” unless the comparison is designed to support that claim and includes the relevant costs.
Data and workload
Document how data is cleaned, compressed, sampled, and encoded. Data preparation can dominate a small experiment, particularly when a classical dataset must be mapped to a limited number of qubits. The team should state whether the data is synthetic, historical, sampled, or representative, and identify any assumptions that would change at production scale.
For encoding details, see the guide to basis, angle, amplitude, and other quantum data encoding methods. For machine learning projects, record train-test separation, feature preparation, model selection, and classical preprocessing as part of the full workflow; the quantum machine learning data pipeline guide provides a useful checklist.
Circuit and hardware evidence
Capture logical qubit count, circuit depth, two-qubit gate count, connectivity assumptions, transpilation or compilation changes, shot count, measurement strategy, queue or access conditions, and the observed error or failure modes. Separate simulator results from processor results. If error mitigation is used, record its sampling overhead, calibration requirements, and effect on uncertainty rather than reporting only the corrected output. Relevant methods include readout correction, zero-noise extrapolation, probabilistic error cancellation, and symmetry verification; compare them using the same workload and reporting format.
Evaluate a vendor or platform on workload fit rather than headline specifications. Useful criteria include access stability, supported software frameworks, compilation transparency, device selection, data handling, experiment controls, observability, documentation, support model, and portability. A platform may be suitable for learning or prototyping without being suitable for a controlled enterprise evaluation.
Reproducibility, security, and economics
Store source code, dependency versions, circuit definitions, compiler settings, backend identifiers, calibration context where available, random seeds, shot counts, raw measurement data, and analysis notebooks. A result that cannot be rerun or audited is weak evidence for a procurement or investment decision.
Include security and governance questions even when the POC is not production-bound. Identify whether sensitive data reaches an external service, how credentials are managed, which users can submit jobs, and what information is retained. A quantum POC does not replace a post-quantum cryptography program; keep those tracks distinct and consult the post-quantum readiness checklist when the experiment involves protected information.
Estimate total effort, not just platform charges. Include problem formulation, data preparation, software development, queue or execution time, repeated runs, error mitigation, classical postprocessing, environment management, security review, and stakeholder time. Label unknowns as assumptions and record a range where precision is not justified.
Worked examples
Example 1: A platform comparison ends in “repeat”
An enterprise team tests the same small optimization formulation on two available platforms. The problem definition scores 4, the classical baseline 2, data preparation 3, circuit performance 3, hardware and vendor fit 4, reproducibility 2, economics 3, and business relevance 3. Using the suggested weights, the overall score is 61. The result is not sufficient for expansion because the classical baseline and reproducibility gates are below 3.
The correct decision is not necessarily “the platforms failed.” The next action is narrower: improve the baseline, package the experiment for reruns, capture compiler and backend settings, and repeat on representative instances. If the revised test still produces no meaningful evidence, the team has a defensible basis for closing the POC.
Example 2: Technical success, weak business case
A variational algorithm runs on a simulator and a processor, and the output distribution is reproducible within the stated uncertainty. Circuit performance scores 4 and reproducibility scores 4. However, data encoding and classical postprocessing require more effort than expected, while the baseline already solves the tested instances adequately. The scorecard produces a technically positive result but a business relevance score of 2.
This POC may be worth expanding as a capability or learning exercise, but it should not be presented as evidence of commercial advantage. The recommended action is to test larger or more representative instances only if a specific business threshold can be identified. The team should also compare the complete hybrid workflow, not just processor execution time. For algorithm context, review the guide to VQE, QAOA, and hybrid optimization workflows.
Example 3: Hardware evidence supports expansion
A team establishes a documented baseline, runs the same workload across repeated sessions, records circuit and backend metadata, and identifies a platform whose connectivity and software tools fit the next workload. The POC does not demonstrate quantum advantage, but it meets its stated objective: reducing uncertainty about implementation and vendor suitability. The overall score is 82, with no category below 3.
An expansion decision should still be bounded. Approve a larger experiment with a defined instance range, budget ceiling, success threshold, data policy, and review date. Do not convert a strong POC score into an unconditional production commitment.
When to recalculate
Recalculate the scorecard whenever an input that influenced the decision changes. At minimum, revisit it when a vendor changes access terms, hardware availability, compiler behavior, supported frameworks, or relevant device characteristics. Revisit the economics when execution volume, engineering effort, cloud rates, or internal labor assumptions change. Revisit the technical score when the circuit, data scale, error-mitigation method, or classical baseline changes.
Use a versioned record rather than overwriting the old assessment. Keep the date, experiment identifier, platform, software environment, input set, score by category, assumptions, and decision owner. This makes it possible to distinguish genuine progress from a change in test conditions.
Before the next review, take these actions:
- Restate the decision the POC is meant to support.
- Freeze the workload, baseline, and scoring rules for the comparison window.
- Run the experiment with raw outputs and execution metadata preserved.
- Score technical evidence separately from business relevance.
- Identify every category below the local gate and assign an owner.
- Choose one outcome—close, repeat, expand, or production-readiness review—with a written reason.
- Set the next recalculation trigger, date, budget boundary, and evidence threshold.
This process keeps quantum hardware and vendor analysis grounded in the actual workload. It also gives leadership a more useful answer than a qubit count or a successful circuit demo: what was learned, how reliable the evidence is, what remains uncertain, and what the organization should do next.