Skip to main content
Q-Sure

Quantum learning with an evidence gate

Know when a quantum result is ready to use.

Q-Sure chooses measurement strategies, allocates state preparations, quantifies prediction uncertainty, is designed to monitor drift, and abstains when the evidence is not strong enough.

Open the public demo Review the evidence

Assurance decision

Evidence supports use

Run

Another strategy is stronger

Reroute

Uncertainty remains too high

Abstain

A valid no-run or insufficient-evidence result is a successful assurance decision.

The assurance layer

Make the decision—not only the circuit—reproducible.

Quantum workflows can fail for reasons that an accuracy score does not reveal: finite measurements, drift, unstable calibration, compilation overhead, or a stronger conventional alternative.

Q-Sure evaluates those conditions together and keeps the acquisition budget, uncertainty criteria, and selected route visible.

A valid no-run, reroute, or insufficient-evidence result is a successful assurance decision.

  1. 01

    Prepare

    Define states, target properties, data splits, reliability gates, and the resource ledger before evaluation.

  2. 02

    Compare

    Evaluate quantum and conventional measurement strategies against strong classical and measured-feature controls.

  3. 03

    Calibrate

    Turn finite-shot predictions into calibrated sets, drift signals, and explicit evidence thresholds.

  4. 04

    Decide

    Run, measure more, reroute, or abstain while recording the evidence and counted state preparations.

Primary demonstration

Quantum-data property learning with a real exit condition.

The frozen simulator study constructs 112 distinct four-site XY-chain states, classifies three interior phases and identifies a phase-boundary class, and predicts three incompatible signed observables plus derived energy density from finite-shot measurements.

Frozen comparison

Held-out quantum-data learning

112 states · 30 streams

Measured-feature control

macro-F1
0.898

preparations
2,688

Strongest post-hoc control

Q-Sure-selected policy

macro-F1
0.870

preparations
1,344

Passed selection gates

Finite-shot quantum kernel

macro-F1
0.817

preparations
543,744

1 / 30 streams passed every gate

A finite-shot quantum fidelity kernel is evaluated beside measured-feature controls under the same frozen data split and preparation ledger.

Candidate quantum path

The kernel learned, but did not pass the reliability gates.

The finite-shot fidelity kernel learned above the four-class prior, yet only one of 30 rerun streams passed every reliability gate and the method did not beat the strongest measured-feature control.

0.817

finite-shot kernel macro-F1

0.935

empirical prediction-set coverage

1 / 30

streams passing every reliability gate

Q-Sure decision

Reroute to the stronger measured strategy.

Q-Sure selected the passing single-copy U-statistic plus RBF policy at 384 preparations per state, or 1,344 amortized preparations per held-out state.

0.870

selected-policy macro-F1

1.000

selective accuracy

0.921

phase-boundary abstention

Demonstrated assurance value

Stop spending shots after the decision is clear.

Across 3,600 paired simulator trials, a sequential Q-Sure policy was compared with a fixed 2,048-shot policy on twelve independently generated basis-and-column problems.

82.9%

mean simulated shot savings

94.8%

decisions resolved

0 / 3,600

incorrect run or no-run decisions observed

This is a sampling-policy advantage under the stated simulator contract. It is not quantum computational, hardware, wall-time, or cost advantage.

Assurance ledger

Fixed policy

2,048

shots per trial

Q-Sure median

128

median shots across all trials

Controlled simulator sampling-policy advantage

One-sided 95% lower bound on shot savings: 69.5%. Incorrect-decision upper bound: 0.083%.

Quantum advantage: not demonstrated

The NMF audit remains a negative control; hardware, scaling, wall-time, and cost gates are untested.

Simulation evidence includes confidence bounds and a separate negative quantum-advantage audit.

NMF control

A useful negative result.

The scaled binary-NMF portfolio tested ranks 4, 6, 8, 10, 12, and 16 with 30 independent bases and three held-out columns per basis.

Exact enumeration, branch-and-bound, rounded nonnegative least squares with local repair, and simulated annealing reached the exact optimum on all 540 evaluated columns.

The present synthetic NMF family is too easy for a credible quantum-advantage demonstration.

Cross-architecture contract

Separate scientific definitions from execution.

  • Offline simulators: PennyLane and Cirq independently implement the retained QAOA control without provider credentials or hardware.
  • Hardware plans: IBM gate-model and D-Wave annealing manifests are explicit not-run artifacts; no hardware result is claimed.
  • Auditable resources: compilation, parameter training, embeddings, and shared matrix work are assigned to explicit ledgers.
Review the open repository Explore EnviTrace quantum AI

Explore EnviCloud

Inspect the assurance decision.

Use the public Q-Sure demonstration to compare evidence, measurement budgets, uncertainty, rerouting, and abstention.

Browse all demos (opens in a new tab)

A claim is an output

Build the evidence before choosing the label.

Talk with EnviTrace about a quantum-data learning problem, a hardware validation plan, or an uncertainty-aware ML workflow.

Discuss a Q-Sure study