The figures we stand behind, and their limits
Each result below states what was measured, under which conditions and what it does not show. When a result stops holding, we withdraw it and say so here.
- Benchmarks
- Preregistered and hash-pinned
- Evidence
- Sealed under a SHA-256 Merkle root
- Labels
- execution_mode on every result
In numbers
Results and capacity
0.971
Median SWAP ratio against SABRE
Source · 630 paired circuits, synthetic topologies, p = 1.79e-10
0.6924
QAOA approximation ratio
Source · Depth-1 MaxCut on 3-regular graphs, simulator
24 qubits
State-vector ceiling
Source · Larger requests refused before allocation
140
MCP tools for AI agents
Source · softqcos-mcp
293
Documented API paths
Source · QCOS OpenAPI contract
180,000
GPU-hours allocated
Source · EuroHPC AI Factory Large Scale allocation EHPC-AIF-2025LS13-019, LUMI-G
Results
What each result shows, and what it does not
| Result | Conditions | What it does not show |
|---|---|---|
| VQE reaches chemical accuracy | Two-qubit molecular models, simulator | Larger molecules, or the same on hardware |
| VQE improves on Hartree-Fock for H4 | Simulator | Accuracy on larger systems |
| QAOA reproduces the 0.6924 reference ratio | Depth-1 MaxCut on 3-regular graphs, simulator | An advantage over classical solvers |
| Error-aware routing, median SWAP ratio 0.971 | 630 paired circuits on synthetic topologies; IQR 0.885 to 1.026; 282 wins, 166 losses, 182 ties; p = 1.79e-10 | The gain on real device topologies, not yet measured |
| Surface codes at distance 3 and 5 decode with BP, OSD and BP+OSD | Simulated noise | Error correction on hardware |
| Shor's algorithm factors 15 | Simulator | Any threat to RSA keys in use |
All results above come from simulators. Grover search, the quantum Fourier transform, phase estimation and quantum machine learning are implemented and will be listed when they pass preregistered tests.
How we test
The rules every figure has to pass
- Preregistered
- The test and its pass rule are written before the runA result that fails its own rule is not published.
- Hash-pinned
- Inputs and code are fixed by hashThe run that produced a figure can be identified exactly.
- Labelled
- Every result carries execution_modeYou can always tell a simulator result from a hardware result.
- Sealed
- Evidence under a SHA-256 Merkle rootChecked offline with softquantus-verify. See verifiable computation.
- Withdrawn in public
- An earlier optimiser benchmark was withdrawnIts setup started the optimiser near the answer. With random starting points its advantage over a published optimiser of the same kind disappeared, so we withdrew it on 25 August 2026 rather than restate it.
Questions
About these results
Why are there so few results?
A figure is listed only when its evidence passes the rules above. Implemented algorithms without published results are named as such.
Why do the results come from simulators?
Hardware access through Softquantus Cloud is in development. Until it is in production, results are labelled as simulator results.
Where is the evidence?
Evidence bundles are in the evidence explorer, and the methods are on the benchmarks page.
Test a result before you trust it
Ask for the evidence behind any figure on this site. If we cannot show it, we will remove the figure.