Essays / Foundations

How to read a quantum claim

A procedure for taking any quantum announcement apart: find the number, find the comparison, find the conditions, and find out who checked. Most claims fail before the third step.

By Christian Stark · Published 8 September 2026 · Reviewed 8 September 2026 · Intro · Living essay

You do not need to understand the physics to evaluate most quantum announcements. Most of them fail on structure long before physics becomes relevant. What follows is a procedure that takes a few minutes and disposes of the large majority of what gets published.

One: find the number

Read the announcement and write down every quantitative claim in it. Not the adjectives — the numbers, with their units.

A surprising fraction of quantum coverage survives this step with nothing written down. Announcements about partnerships, national strategies, funding rounds, roadmaps, and future availability contain no measurement. They may be commercially significant. They are not evidence about a machine, and they cannot be checked.

If there is no number, stop. You have finished reading.

If there is a number, note which kind it is. Qubit count is the weakest, and it is the one most likely to appear alone. Gate fidelity, coherence time, logical error rate, and circuit depth are stronger, because each of them constrains what the device can actually do. A press release that gives a count and omits the fidelity has chosen which of those to tell you.

Two: find the comparison

A number without a baseline means nothing. Faster than what? Better than which alternative, run by whom, with how much effort?

This is where advantage claims usually come apart. The claim is that a quantum processor did something a classical computer could not do in reasonable time. The strength of that claim rests entirely on the classical side of the comparison, and the classical side is generally computed by the same team, using the best classical algorithm they were aware of at the time.

That is not dishonest, but it is fragile. The history of this field is that classical algorithms improve after the announcement, often sharply, once specialists have a concrete target to attack. Simulations that were said to require millennia have been brought down to days by better tensor-network methods and more careful use of memory. The physical experiment stays valid; the comparison it was measured against does not.

So the question is not “did the quantum device do it” but “what exactly was the classical baseline, who produced it, and has anyone hostile to the claim tried to beat it.” If the answer to the last part is no, the claim is provisional regardless of how careful the experiment was.

Three: find the conditions

Every result was obtained under conditions somebody chose. List them.

Was the problem selected because it suits the hardware? Random circuit sampling and similar tasks are chosen precisely because they are hard classically and natural quantumly. That makes them good tests of a device and poor evidence about applications. Nobody has this problem.

Was there post-selection? Discarding runs that failed a check is a legitimate technique, and it also means the reported fidelity is conditional on a filter that a real computation would not be allowed to apply.

How was the answer verified? There are three cases, and they carry very different weight. The problem was classically simulable, so the answer could be checked directly — reassuring about correctness, but then a classical computer solved it too. The answer was checked statistically, by arguing that the output distribution has the right shape — weaker, and dependent on the model used. Or the answer was not verified at all, and the claim rests on the machine having behaved as expected.

How many times did it run, and what happened on the runs that were not reported? A result that holds on one device on one day, and has not been repeated, is a data point rather than a finding.

Four: find out who checked

Look at where the claim was published and who has examined it.

A preprint is the primary source and is usually the best available document; it also has not been refereed. A peer-reviewed paper has passed a small number of anonymous readers, which is a real filter but not a reproduction. A company blog post describing internal work with no accompanying paper is an assertion.

The strongest signal, and the rarest, is an independent group reproducing the result or failing to. Reproduction in this field is expensive, because the apparatus is expensive, so the honest position on almost every recent result is that it has not been independently confirmed.

Refutations deserve the same scrutiny as claims. A classical algorithm that matches a quantum result is itself a result, and it should be read with the same four questions.

What survives

Applied consistently, this procedure discards most of what is published about quantum computing. That is the correct outcome. The field produces a handful of results a year that change what is known, and a very large volume of material that changes only what is being said.

Three sentences summarise the whole procedure, and they are worth keeping somewhere near the front of your mind.

A press release is not evidence. It is a description of evidence, written by a party with an interest in how it is described, and the primary source is generally available and generally more careful.

Laboratory advantage is not application. Doing something faster than a classical computer on a task designed to be hard for classical computers is a measurement of the device, not a use for it.

A demo is not a machine. The demonstration establishes that a thing happened once, under chosen conditions, checked by the people who arranged them. That is how progress is made, and it is not the same as a computer somebody can use.

None of this requires assuming bad faith. Researchers in this field are, on the whole, careful about caveats, and the caveats are usually right there in the abstract. They are removed in transit — by summaries, by headlines, by the need to describe a slow technical process as an event. Reading the primary source puts them back.

This is a foundation essay. It is kept current rather than frozen, so the reviewed date above matters more than the publication date.