How KetQat compares to other estimators

KetQat's figures are checked against Microsoft's QDK and Google's Qualtran. Where the tools disagree, the cause is named and both values are shown. Nothing is averaged.

Agreed: 5
Differed: 1
Could not run: 0
Observed: 2026-08-15

Where they disagree, and why it is not a bug

required code distance — KetQat 13, qualtran 15.

The logical-error prefactor is fitted, not derived. This project defaults to A = 0.03 (Fowler conventional); Qualtran's make_gidney_fowler uses A = 0.1. Neither is more correct, and the resulting distances differ by one step. KetQat reports both as model sensitivity on every estimate rather than choosing one and presenting it as the answer.

Every comparison in this run

Every quantity compared against a reference estimator, with both values and the outcome. No aggregate row is produced: the quantities are different quantities, so a combined figure would describe nothing.
Quantity	Against	KetQat	Reference	Outcome
logical qubits including routing space2n + ceil(sqrt(8n)) + 1 at n=4. Beverland et al. (2022), the formula Microsoft's estimator implements.	qdk	15	15	AGREED
logical qubits including routing space2n + ceil(sqrt(8n)) + 1 at n=8. Beverland et al. (2022), the formula Microsoft's estimator implements.	qdk	25	25	AGREED
logical qubits including routing space2n + ceil(sqrt(8n)) + 1 at n=16. Beverland et al. (2022), the formula Microsoft's estimator implements.	qdk	45	45	AGREED
logical qubits including routing space2n + ceil(sqrt(8n)) + 1 at n=32. Beverland et al. (2022), the formula Microsoft's estimator implements.	qdk	81	81	AGREED
logical qubits including routing space2n + ceil(sqrt(8n)) + 1 at n=100. Beverland et al. (2022), the formula Microsoft's estimator implements.	qdk	230	230	AGREED
required code distanceQualtran importable; its documented default prefactor is 0.1.	qualtran	13	15	DIFFERED

How to read this

Could not run is not a pass. A comparison whose reference tool was unavailable proves nothing about either estimator, so it is reported as its own outcome. A suite that silently skipped it would report a clean sweep it had not earned, and the CI gate fails when a required comparison is unavailable.

Differed is not necessarily a defect. The logical-error prefactor is fitted, not derived. Choosing 0.03 or 0.1 is a modelling decision, and two tools choosing differently is a definitional difference — recorded here, and carried as sensitivity on every estimate, rather than tuned away until the numbers match.

Agreement is not validation. Two models agreeing is evidence that they agree. It is not evidence that either describes a machine that exists, and this product never presents it as such.

This is a dated observation. It records what those tools said on 2026-08-15 at Microsoft QDK (qdk) 1.31.0 and Google Qualtran 0.7.0. It is superseded by a newer run, never edited in place. Reproduce it with python -m ketqat_benchmarks.resource_intelligence --require qdk qualtran in ketqat-benchmarks.

Read the model these figures come from or what KetQat will not tell you.