August 16, 2026
Eight numbers beat a neural network on quantum weak‑measurement records
The finding
A linear model on 8 time-binned means scores 0.820 held-out accuracy; an MLP on the raw 64-sample record scores 0.805. Naive summary statistics fail only for lack of temporal resolution — crude binning recovers everything, and the raw series adds nothing.
We went looking for a case where a learned model on a raw time series would beat summary statistics, because our own conclusion depended on finding one. We chose weak-measurement records deliberately: a continuous readout trace is the most structural, least obviously summarisable input this domain offers. If the claim was going to break anywhere, it was there.
It did not break.
The ladder
Held-out accuracy on the same discrimination task, four ways:
| discriminator | held-out accuracy |
|---|---|
| linear on the record mean (1 number) | 0.570 |
| linear on 8 signal-quality statistics | 0.574 |
| linear on 8 time-binned means | 0.820 |
| MLP on the raw 64-sample record | 0.805 |
Read the ladder rather than the headline. One number is nearly useless. Eight carefully chosen signal-quality statistics are no better — 0.574 against 0.570, which is the result that surprised us. Eight crude time bins jump to 0.820. And the full 64-sample series, handed to a network, comes back slightly worse than the eight bins.
So the binding constraint was never dimensionality. It was temporal resolution — and the cheapest possible temporal summary captures essentially all of it.
Why this is structural, not a quirk of our setup
A Lindblad master equation is linear in its dissipators. Figures of merit are therefore near-separable in the noise rates, and a continuous weak measurement of a few-level system cannot produce an un-summarisable record: its sufficient statistics are low-dimensional by construction.
That is a statement about the physics, not about our data. It generalises to any effective-rate model, which is the reason we stopped rather than trying the same experiment in a neighbouring domain — a plasma-control parameter space has the same anchors and would give the same answer.
What we are claiming, and what we are not
Not that networks are useless on quantum data. The claim is narrower and more useful: before reaching for a model on a raw record, check whether the record is summarisable. For weak measurement of a few-level system, the answer is yes, and it is provable rather than empirical.
Honest provenance: this came out of an experiment whose pre-registered question — does device drift harm readout enough for a detector to help — was adjudicated INCONCLUSIVE, because drift turned out not to harm readout at all. The summarisability result is from a detector-free test run alongside it. It is a side result of an inconclusive experiment, and we would rather say so than present it as the thing we set out to find.