Reality Board
Four questions. Not forty. Each one is a thing that either has happened or has not, and each has a stated condition for changing the answer.
The board exists so that a reader can check the state of the field without reading a year of announcements. Most of these cells will read Not yet for a long time. That is information, not pessimism.
A cell moves only when a human reviews the evidence against the condition written under it. Automated tools may flag a candidate for review. They cannot change a cell.
-
Early Reviewed 24 September 2026
Stable, repeated error correction below threshold
Below threshold means that making an error-correcting code larger makes the encoded information better, not worse. It is the sign change that decides whether any of this scales.
There is now published work in which increasing the code distance lowered the logical error rate, on real hardware, over repeated correction cycles. That is a genuine result and it is why this cell reads early rather than not yet.
What it is not is a solved problem. The demonstrations are short relative to the duration of a real computation, they run on a small number of logical qubits, and the margin below threshold is thin enough that the overhead required for a useful logical error rate remains very large. Sustained operation over the timescale of an actual algorithm has not been shown.
Changes to this answer
- 2026-09-24 — Raised 20 September, reviewed 24 September, not applied. The published below-threshold memory (arXiv:2408.13687) decodes in real time for a million cycles at a 1.1 microsecond cycle time, which is about one second of operation on one platform. The condition below asks for hours and for a second platform, so the answer stays Early.
What would change this answer
- Below-threshold operation sustained over hours rather than seconds, with the logical error rate stable throughout.
- The same scaling reproduced on a second hardware platform by a group with no stake in the first result.
- A margin below threshold wide enough that the physical-to-logical overhead falls substantially rather than marginally.
-
Early Reviewed 8 September 2026
Logical qubits with useful operations
Encoding a logical qubit is not the same as computing with one. This cell asks whether logical qubits can be entangled, operated on, and kept correct while a sequence of logical gates runs.
Small numbers of logical qubits have been prepared and operated on, and entangling operations between them have been demonstrated. The logical error rates involved are still high, the circuits are short, and the number of logical qubits is in the tens at most.
The hard part is universality. Clifford operations alone are classically simulable, so a machine restricted to them cannot outperform a laptop. Getting past that requires non-Clifford gates, which in most architectures means magic state distillation, and distillation is where the physical qubit budget goes. No system has demonstrated a universal logical gate set running under continuous error correction at a useful error rate.
What would change this answer
- A universal logical gate set, including non-Clifford operations, running with error correction active throughout.
- Logical error rates that stay flat across a deep sequence of logical gates rather than degrading with circuit depth.
- Magic state distillation demonstrated at a cost consistent with published resource estimates rather than far above them.
-
Not yet Reviewed 8 September 2026
Advantage on a scientific or industrial problem someone actually has
The qualifier is the whole cell. Advantage has been claimed on sampling tasks chosen because they are hard for classical computers, and several of those claims have since been narrowed by improved classical methods. Nobody needs the output of a random circuit.
This cell asks a different question: has a quantum processor produced a result that somebody wanted for reasons unconnected to quantum computing, and that could not be obtained classically? A ground state energy at chemical accuracy for a molecule beyond classical reach. A correlation function a materials group could not otherwise get. An optimisation result a business would pay for.
The answer is no. Variational methods on current hardware remain limited by noise and by measurement cost, and where they produce a correct answer, a classical method generally produces it too. Laboratory advantage is not application.
What would change this answer
- A result published by domain scientists — chemists, materials scientists, engineers — who wanted the answer for their own work.
- A classical community given the problem and the data, failing to match the result after a serious attempt.
- Reproduction of the result on different hardware, since a one-off on a single machine is a data point rather than a finding.
-
Not yet Reviewed 8 September 2026
Fault-tolerant machine at useful scale
A machine that runs an arbitrary algorithm of the user’s choosing, long enough to finish, with error correction holding for the whole run, at a scale where the algorithm does something worth doing.
This is the end state. Published resource estimates for the canonical applications — factoring at cryptographic scale, simulating industrially relevant chemistry — sit in the range of millions of physical qubits with sustained correction. Current devices are three to four orders of magnitude short, and the gap is not only in count: it is in error rate, in connectivity, in classical decoding throughput, and in the cryogenic and control engineering required to run that many channels at once.
This cell will read not yet for a long time. Watching it change is not the point; watching cells one and two is. This one moves last.
What would change this answer
- A machine executing a user-chosen algorithm of real depth, with error correction active for the full run and a verifiable answer.
- Resource estimates and delivered hardware converging, rather than estimates falling while hardware stays flat.
- An outside party running their own workload on the machine and getting a correct answer they did not know in advance.
What is deliberately not on this board
There is no cell for cryptographic threat. The question of when a cryptographically relevant quantum computer arrives is downstream of cells two and four, and giving it its own row would suggest it moves on its own. It does not. It moves when the hardware moves.
There is also no cell for funding, headcount, partnerships, or share price. None of those are evidence about a machine.