Discarding most of an experiment’s results sounds like a way to manufacture a better answer. In a quantum test, it can instead be a legitimate part of the question, provided the selection rule is fixed in advance and the comparison is rebuilt around that same rule.
A paper published in Quantum on November 5, 2026, explains how to do that using what its authors call “post-selection games.” The useful idea is not permission to hide disappointing measurements. It is that a test can count some outcomes and set others aside without giving quantum behavior an unfair advantage over the classical explanation it is being tested against.
The paper’s concrete example uses data from a 2015 experiment. The new work organizes and extends a way of analyzing evidence; it does not report a newly built quantum machine.
The Comparison Must Survive the Filter
The setting is a Bell test, named after physicist John Bell. Such experiments ask whether correlations between separated measurements can be explained by a local model: each side produces its answer using information available on that side, without receiving the other side’s current measurement choice.
A useful way to picture the test is as a game between two separated players. Each receives one of two questions and gives one of two answers. They can agree on a plan beforehand, but cannot communicate during a round. The referee then checks the questions and answers against a fixed scoring rule.
Quantum systems can produce correlations that beat the best local strategy in certain games. That is a statement about the pattern of measurement results, not a way to send messages faster than light. It is also a different question from whether a quantum computer can finish a useful commercial task.
Now add a third possible verdict: the referee can discard a round rather than call it a win or loss. This is post-selection, meaning selection after the outcomes are available. The selection rule itself is part of the test, not something invented afterward to keep only a pleasing result.
The crucial move is to ask how well a local strategy could score after the same filter. Comparing a filtered quantum score with a classical limit calculated for a different, unfiltered game would not answer that question.
A Fair Filter Still Leaves Losing Outcomes
In the paper’s example, based on Hardy’s paradox, the filter retains four particular combinations of questions and answers. One is a win; the other three are losses. All other combinations are discarded. A rule that kept only wins would be useless: a local strategy would look perfect under that rule too.
With the question pairs equally likely in the model, the retained-round winning probability for a local strategy cannot exceed 50%. The reason is concrete. Any fixed response plan that can produce the winning combination also produces at least one of the retained losing combinations.
There are only 16 fixed response plans in this two-question, two-answer example. Enumerating them reproduces the 50% bound. Randomly mixing those plans cannot improve it: the retained results are still a weighted mixture of plans whose expected winning rates are at most one half.
That is a bound on the expected probability, not a promise that every finite sample from a local process will stay below 50%. Random fluctuations still happen. A real experiment needs a statistical analysis, not merely a score on the favorable side of a line.
What the 2015 Measurements Actually Show
The new paper illustrates the game using a dataset from a photon experiment reported by Lynden Shalm and colleagues in 2015. The selected dataset contains 177,358,351 original rounds. The filter retains 12,127 of them, including 6,378 wins: about 52.6% of the retained rounds.
Those counts can be traced back to the original experiment’s supplementary table. The meaningful comparison is with the selection-adjusted 50% local bound, under the game’s assumptions. It is not a comparison with the 75% bound of the familiar, differently scored CHSH game.
But 52.6% alone is not a complete verdict on the experiment. How the settings were chosen, how the test handled uncertainty, and the stopping rule used to define the analyzed sample also matter. The original researchers collected data for fixed periods, then applied a predefined retained-event stopping rule to the analysis. They explicitly examined how choosing a stopping point after seeing the results could exaggerate the evidence.
There is another important historical limit: the 2015 analysis already used this statistical selection idea. The 2026 paper acknowledges that precedent. It would be wrong to say the new authors rescued an old experiment that had failed to establish its result.
Their contribution is to treat selection as part of the game itself, with its own calculable classical and quantum limits, and explore the resulting tests. The simple example makes that contribution understandable without treating every wider mathematical claim in the paper as an experimentally demonstrated advantage.
The Discarded Rounds Still Have a Cost
Only about 0.0068% of the original rounds survive this filter. A high score among those survivors cannot tell us, on its own, how efficiently the experiment gathered evidence. Producing the discarded rounds still took experimental effort.
The paper accounts for the probability that a round is retained when comparing statistical information per original round. That denominator matters. A method that produces an impressive result from a tiny selected subset must still be judged against the work needed to obtain that subset.
Nor does throwing away data always help. In one idealized example in the paper, the conventional comparison is better. The appropriate test depends on the behavior and imperfections being measured. There is no general rule that a smaller dataset means a stronger conclusion, and no quantum-computer speedup follows from this analysis.
For the broader distinction between a convincing quantum experiment and useful computing, what makes a quantum computer useful is the next question. Bell-test evidence establishes a different kind of result, a distinction that also runs through Alain Aspect’s move from quantum foundations toward applications.
A selected score is worth taking seriously when the selection rule is visible, the competing explanation faces that same rule, and the uncertainty and discarded effort remain in view. The problem is not that some measurements were set aside. It is whether the comparison quietly changed with them.
