Suppose you want to pay an untrusted company to run a quantum computation for you, and you want a public smart contract to release the payment only if the work was really done on a quantum device. A smart contract is a program on a public blockchain whose every operation is metered and paid for in "gas." Protocols already exist that let a classical verifier check quantumness cheaply. These proofs of quantumness are challenge-response games in which an honest quantum device scores above a bound that no classical machine can beat. The difficulty is that real hardware is noisy. Noise pushes the device's score toward the classical bound, and the closer the score gets to that bound, the more measurement samples the verifier must check before it can tell quantum from classical. Each checked sample costs gas.
This note argues that the cost of settling quantum work on-chain is set mainly by one hardware number, the margin between the device's expected score and the classical bound. Settlement gas grows as the inverse square of that margin, and it diverges at a threshold hardware fidelity. We propose to build the contract, measure the relationship, and use it in two ways. First, the contract checks samples one batch at a time and stops as soon as the evidence is sufficient, so its sample count adapts to day-to-day hardware drift. Second, the provider picks circuit placements on the chip to maximize margin rather than to minimize generic output error. A further output is a measured crossover map: at which margins direct on-chain checking, a zero-knowledge proof of the check, or an optimistic challenge window is the cheapest way to settle.
Core Claim
For a proof-of-quantumness protocol with per-sample score variance $\sigma^2$, the number of samples needed to reject the classical hypothesis at confidence level $z$ (in standard deviations) is approximately
$$m \approx \left(\frac{z\,\sigma}{\delta}\right)^2, \qquad \delta = F\,(s_{\text{ideal}} - s_{\text{rand}}) - (b - s_{\text{rand}}).$$
The first expression is the normal-approximation sample size for a one-sided mean test. In the second, $F$ is an effective circuit fidelity, $s_{\text{ideal}}$ the noiseless score, $s_{\text{rand}}$ the score of uniformly random outputs, and $b$ the classical bound. The second expression assumes noise mixes the ideal output with random output. Settlement gas is roughly $g_0 + g_1 m$, where $g_0$ is the fixed cost of commitments and payout and $g_1$ the gas to check one sample. Gas therefore scales as $1/\delta^2$ and blows up as $F$ approaches the threshold $F^\ast = (b - s_{\text{rand}})/(s_{\text{ideal}} - s_{\text{rand}})$.
The prediction has a concrete anchor in the file. Interactive Protocols for Classically-Verifiable Quantum Advantage ran LWE- and factoring-based proofs of quantumness on trapped ions and cleared the classical thresholds at more than 6σ (LWE) and more than 4σ (factoring, moduli 8 to 16). Its deepest instance, N=21, fell just below the threshold because of accumulated gate error. In our terms, that instance sits past $F^\ast$, where no amount of gas suffices. The claim that is new here is that noise mitigation and qubit placement should be evaluated by margin gained per unit of gas, not by fidelity in the abstract.
Relation to Prior Work
Two papers supply the verification protocols. Zhu et al. (Interactive Protocols) give trapdoor-based interactive rounds checked on a local computer. Instantaneous Quantum Polynomial-Time Sampling and Verifiable Quantum Advantage gives an IQP sampling scheme, meaning circuits built from commuting gates, verified through correlations with a hidden code. That paper is analyzed only for classical hardness, with no physical noise. Neither paper considers a public verifier, which raises a problem they never face: on-chain everything is public, including the trapdoor.
Two papers supply hardware-noise methods we reuse and change. Anchor shows that identical circuits on 127-qubit IBM machines swing between 10% and 40% total variation distance (TVD), a measure of how far the output distribution is from the ideal one. Anchor stabilizes this by predicting TVD with a regressor and spreading shots across many qubit placements. We reuse its placement pool but change the objective from TVD to margin. QASMBench reports that circuits with uneven two-qubit interaction patterns are very sensitive to placement, which is why placement is the lever we target.
Several papers motivate and constrain the blockchain side. Decentralized Anonymous Micropayments cuts on-chain frequency 1,000-fold with lottery tickets, an idea we treat as a complement rather than a substitute. Impact of EIP-4844 describes cheap temporary data "blobs," which we use for committed samples. The research file also notes that verifying succinct cryptographic proofs on-chain carries a nontrivial gas floor, and that floor determines where direct checking can win.
The file's own survey found no work that settles quantum proofs on a public ledger. Quantum-blockchain papers such as QBIoT and SoK: Blockchain Consensus in the Quantum Age either assume ideal quantum channels or treat quantum computers only as attackers.
Research Program
Contract design. We will build a settlement contract around a commit, challenge, reveal sequence that needs a fixed number of transactions regardless of shot count. The requester generates fresh keys for each job, publishes the public parameters, and posts only a hash of the trapdoor. The trapdoor is the factorization for the Rabin protocol, the LWE secret, or the hidden code vector for the IQP scheme. The prover runs the first-round circuits for all shots and posts a Merkle root of the first-round outputs. A Merkle root is a single 32-byte hash committing to the whole list, and the outputs themselves go in a blob. The contract then derives every per-shot challenge from post-commitment randomness. The prover runs the second round and commits those answers. Only then does the requester reveal the trapdoor, so it is useless for cheating on this job. If the requester refuses to reveal, it forfeits a deposit.
Sequential checking. Verification proceeds in batches. The prover opens Merkle paths for a batch of shots chosen by the contract, the contract checks each one, and it updates a sequential likelihood-ratio test, a stopping rule that halts once evidence crosses a preset confidence. Because the contract stops early when the margin is large, a good hardware day costs less. Because the stopping rule is valid at any sample size, a drifting device cannot be underpaid or overpaid by a sample size fixed in advance.
Margin-driven placement. On the provider side, we will train Anchor's regressor to predict the protocol score from calibration data instead of TVD. We will then choose the placement and shot split that maximizes predicted $\delta^2$ per unit of combined quantum-processor and gas cost.
Deliverables. The project produces the contract, the cost model, the margin predictor, and a measured map of which settlement backend is cheapest at each margin.
Baselines
The naive reference is trusted attestation. A named off-chain verifier checks the proof locally, as in Zhu et al., and posts a signed pass or fail. This has near-zero quantum-specific gas but reintroduces the trusted party that decentralization was meant to remove. We report its gas to fix the floor.
The second baseline is literal on-chain interaction, one transaction per round per shot. It shows the cost of ignoring batching and is expected to be orders of magnitude worse.
The third is batched, fixed-sample verification. It uses our commit, challenge, reveal sequence, but the sample size is set in advance from a worst-case margin. This isolates the value of sequential stopping.
The strongest competitors are two alternative backends run on the same committed data:
- Proof-of-verification backend. After the trapdoor is revealed, the prover produces a succinct zero-knowledge proof that the classical check passed on all shots. The contract verifies only this proof at near-constant gas.
- Optimistic dispute backend. The claim pays out after a challenge window unless disputed, and our sequential checker resolves disputes.
All backends run on identical shot data, identical noise profiles, and the same local Ethereum Virtual Machine with fixed gas schedules. Each tuning knob, such as batch size, stopping thresholds, and proof-system parameters, gets the same grid-search budget.
The key ablation compares TVD-optimal placement, which is Anchor as published, against margin-optimal placement on identical device profiles. For the hardware-side baseline, the reported numbers are Anchor's: a 73% average reduction in TVD standard deviation, a 27% average error reduction, and stable 12 to 15% TVD on 127-qubit IBM machines versus 10 to 40% for traditional methods (Anchor).
There is no reported result on Lacuna for the gas cost of verifying any proof of quantumness on-chain. Our baseline numbers will therefore be the first. Beating the baselines means two things. Sequential checking must reduce expected gas relative to fixed-sample checking at equal false-acceptance rate across drifting profiles. Margin-optimal placement must reduce gas at equal confidence relative to TVD-optimal placement.
Why not just…?
Why not just use an optimistic challenge window and never verify on-chain? In the cooperative case this is cheaper than anything we build, and we will report it that way. Its weakness is enforcement. A requester who wants to avoid payment can force a dispute, and the dispute is settled by full verification. If dispute-path gas exceeds the job's value, honest provers are open to griefing. Our checker is meant to be the dispute path. The crossover map will show for which margins the dispute path stays affordable. If the dispute path is affordable at all realistic margins even without sequential stopping, our contribution narrows to the cost model.
Why not just wrap the classical check in a zero-knowledge proof and pay constant gas? This is the most serious threat. The check itself is classical and cheap, so proving it succinctly is feasible, unlike proving a quantum simulation. Noise does not disappear in this design. It moves into the proof generation, whose size grows with $m$, and the on-chain proof verification has a fixed gas floor that small jobs may not clear. We will measure both backends side by side. If proof generation is cheap and its floor is below direct checking at every margin we observe, then the noise-to-gas relationship matters only for off-chain cost, and we will say so.
Why not just audit a random fraction of jobs, lottery-style? Probabilistic auditing in the spirit of Decentralized Anonymous Micropayments divides expected gas by the audit rate. It also needs slashing deposits large enough to deter cheating, and each audit still costs what our model predicts. Auditing is complementary: it scales our per-audit cost rather than replacing it. We will report audited configurations at audit rates of 1/10 and 1/100.
Why not just pick placements by lowest predicted TVD, as Anchor does? TVD weights every output equally, while the protocol score depends only on specific correlations. A placement that is slightly worse in TVD may protect exactly those correlations. The placement ablation tests this directly. If TVD-optimal and margin-optimal placements give indistinguishable gas, the provider-side contribution reduces to using Anchor unchanged.
Proposed Experiments
Workloads and noise. We will use the Rabin-factoring and LWE instances from Zhu et al. (moduli 8 to 21), IQP circuits of the hidden-code type from Bremner et al. with π/8 phase rotations, and QASMBench medium-scale circuits as placement stress tests. Noise comes from calibration profiles of the five IBM devices used in Anchor, replayed across calibration days to create drift. Where hardware access allows, we will also run small instances on real machines.
Gas measurement. We measure gas on a local Ethereum Virtual Machine, with committed samples posted as blobs. The primary metrics are expected settlement gas per job at a fixed false-acceptance rate, the false-acceptance rate itself under a classical cheater who submits the best classical strategy at the bound, and the false-rejection rate for honest provers under drift.
Decisive checks. Four results decide the outcome:
- Measured gas plotted against measured margin should follow $1/\delta^2$ and diverge near $F^\ast$. The N=21 factoring instance should be unsettleable at any budget, matching Zhu et al.
- Sequential checking should hold the honest-rejection rate fixed across Anchor's drifting profiles while fixed-sample checking does not.
- Margin-optimal placement should beat TVD-optimal placement in gas at equal confidence.
- The crossover between direct checking and the proof-of-verification backend should fall at a margin we can locate. Locating it tells practitioners which backend to use for a given device.
If the first check fails, meaning gas does not track margin in this way, the central claim is wrong and the remaining results do not matter.
References
- Interactive Protocols for Classically-Verifiable Quantum Advantage
- Instantaneous Quantum Polynomial-Time Sampling and Verifiable Quantum Advantage: Stabilizer Scheme and Classical Security
- Anchor: Reducing Temporal and Spatial Output Performance Variability on Quantum Computers
- QASMBench: A Low-level QASM Benchmark Suite for NISQ Evaluation and Simulation
- Decentralized Anonymous Micropayments
- Impact of EIP-4844 on Ethereum: Consensus Security, Ethereum Usage, Rollup Transaction Dynamics, and Blob Gas Fee Markets
- QBIoT: A Quantum Blockchain Framework for IoT with an Improved Proof-of-Authority Consensus Algorithm and a Public-Key Quantum Signature
- SoK: Blockchain Consensus in the Quantum Age