Programme · Concluded
Learned readout discrimination
Does a learned model beat the analytic statistic at classifying quantum measurement records?
Why this direction
Readout discrimination looked like the best available fit between machine learning and the superconducting stack: a high-dimensional signal, a classification target, and abundant data. If a learned representation could extract more from a measurement record than the standard filter, it would pay out on every shot of every experiment.
Why it ended
It died twice, in two sessions, for $0 each.
The first kill was that the criterion had already been published and the experiment already run — with a result contradicting the prediction. The second was that the surviving thread rested on a separation that does not exist.
What survived is a reference fact worth more than the programme was: the correct baseline for this work is the trainable temporal postprocessor, not the matched filter. The matched filter is the special case the TPP reduces to exactly under Gaussian white noise, which means benchmarking a network against the matched filter on correlated noise is measuring the wrong gap. That single correction is published below, because it is the part a reader can use.
What would reopen it
A demonstration that some learned statistic beats a properly specified temporal postprocessor on real device data, with the interval reported — not against a matched-filter baseline chosen because it is easy to beat.
Findings
1- Aug 2026
Benchmarking learned readout against the matched filter measures the wrong gap
The trainable temporal postprocessor reduces exactly to the matched filter under Gaussian white noise, so the matched filter is a special case rather than a peer. Beating it on correlated noise demonstrates that the noise was correlated, not that learning helped.