QEC Explorer isn't just a teaching demo. It's built around two questions the quantum-computing field is actively working on right now: how do we build codes that don't waste enormous numbers of qubits, and how do we decode them well enough to matter? This page connects the interactive modules and notebooks to that research, in plain terms.
The surface code you play with in Module 1 is the field's current workhorse, and for good reason: every check it makes only involves neighboring qubits on a flat grid, which is exactly what today's hardware can build. But that convenience has a steep price. To protect a single piece of quantum information well enough to run a real algorithm, a surface code needs hundreds to thousands of physical qubits per logical qubit. Scale that to the millions of physical qubits a useful machine would need, and the overhead becomes the central engineering obstacle.
A newer family, quantum low-density parity-check (qLDPC) codes, keeps the good property that each check still only touches a few qubits, but relaxes the demand that those qubits be physical neighbors. By allowing a small number of longer-range connections, these codes pack many logical qubits into far fewer physical ones. IBM's bivariate-bicycle codes are the headline example: their "gross code" protects 12 logical qubits with 144 data qubits at the same distance where a surface code would protect just one, roughly a tenfold cut in overhead.
The trade-off is real: those longer-range links are harder to wire in hardware, and they make decoding harder, which is exactly where the second research thread comes in.
A decoder is the algorithm that looks at the pattern of lit detectors and guesses what error occurred. It's the puzzle you watch three decoders attempt in Module 2. The catch, which you can see for yourself, is that the syndrome never uniquely determines the error. Many different errors light up the same detectors. Decoding is really a problem of inference under ambiguity, and the best general-purpose classical method, belief propagation, actually fails on the surface code, as Notebook 3 shows.
The frontier response is to learn a decoder from data instead of hand-designing it. Google's AlphaQubit (2024) is the landmark result: a neural network trained on enormous amounts of error data that decodes large surface codes better than the classical standards. Notebook 5 builds a small, real version of this idea, a graph neural network that passes messages over the same detector graph the other decoders use, trains it on synthetic noise, and benchmarks it against belief propagation.
And here's the part worth being upfront about: at the tiny scale a teaching notebook can train, the neural decoder does not beat the classical baseline. Notebook 5 reports that result as it is, and explains why scale is what makes learned decoders win. The gap between a small demo and a state-of-the-art result is the research.
Better codes (fewer qubits) and better decoders (smarter inference) are the two halves of making quantum error correction practical. QEC Explorer walks the full path to that frontier, with code you can read and run behind every step.
Note: specific figures (e.g. surface-code overhead "per logical qubit," the ~10× improvement) are drawn from the sources above; where a notebook uses a round number for illustration rather than a single sourced constant, it says so explicitly. This summary is written in the author's own words.