A quantum computer's memory is guarded by an endless stream of parity checks. When checks light up, something went wrong inside — and someone has to work out what, fast, or the stored information is lost. Below: checks from a simulated honeycomb quantum memory, and a real neural network that learned to read them. It is running live in this page — no server, no tricks.
every number on this page is measured, in your browser, on the nets you're playing againstErrors just struck this memory — the sparks are its alarm bells. Watch them shimmer for a moment, then have a hunch: did the two stored quantum bits (A and B) make it through? No grades, no pressure — honestly, nobody can see this pattern by eye (~28% is what anyone scores), and that's exactly the point of what comes next.
A new burst loads after each reveal. Whenever you've had enough, scroll on — the net plays this game two thousand times in Round 2.
Same game, played fast. The net decodes each burst of sparks live in your browser; the classical gold-standard algorithm (minimum-weight matching, the one real quantum computers use today) answered the same shots ahead of time.
Honesty first: the net does not beat the gold standard here, and we don't claim it does. What's remarkable is sitting one section down.
Here are identical twin networks: same size (63k parameters), same training examples, same everything — except one was hard-wired with the honeycomb's symmetry: "rotate the sparks 120°, and your answer must rotate the same way." The other twin had to figure the task out alone. Run them on the same 300 shots:
That's the research result this page exists to show: on this code, telling the network about the lattice's symmetry isn't a speed-up — it's the difference between learning and not learning. We proved the symmetry exactly from the error model (it even swaps the roles of A and B as it rotates), then built it into the net.
Real machines can't hand you millions of labelled failures. Drag the slider — each stop loads the actual symmetry-aware net trained on that many examples and runs it on 300 shots, live.
For scale: the ordinary twin never beats guessing even with 64,000 examples — the symmetry-aware net is already ahead of that with 250.
— Quantum computers fail constantly; an error-decoder
must call every failure correctly, at speed, forever. It's a real bottleneck in the field.
— Labelled failure data from real hardware is scarce, so AI that learns from few
examples is what matters — not bigger models or more compute.
— The unlock here wasn't horsepower. It was mathematics: proving the memory's exact
symmetry from its error model and building that proof into the network.
| training examples | ordinary net | symmetry-aware net | gold standard |
|---|---|---|---|
| 250 | — | 34.4% | 94.7% |
| 1,000 | — | 40.5% | |
| 4,000 | — | 46.7% | |
| 16,000 | — | 59.4% | |
| 64,000 | 29.1% (≈ guessing) | 74.7% |
Measured on 200,000 held-out shots, on the exact float16 nets shipped with this page (simulated circuit-level noise, one code size, honeycomb Floquet memory at L=6). The gold standard remains unbeaten and we claim no exception. Technical readers: the symmetry is an exact p3 space group of the detector error model with a GL(2,F₂) twist that mixes the logical qubits — solved, verified and shipped via demsym (pip install demsym); preprint in preparation.