Workbench

The workbench is eighteen research tools at /workbench, each a focused page over a tested, dependency-free module in lib/workbench/ or lib/quantum/. The design rule is the same as the simulator's: the computation is a pure module with a unit test against a known result, and the page is a thin view over it.

Tool	Kind	What it does
Generate	generate	Generate → verify → repair loop: a model proposes a circuit against a spec, the verifier checks histogram, state or unitary and resource constraints, and the loop repairs. The provider is injected, so the loop is tested with a scripted fake
Algorithm Evolution	optimise	A demo-scale replica of an evolutionary program search. Candidates are programs in a tiny line-based DSL that is parsed and interpreted; nothing from a model is ever evaluated as code
Decoder Lab, Code Explorer, Stabilizer	measure	Surface-code decoders (lookup versus a learned decoder), Clifford circuits on thousands of qubits via the tableau engine
Calibration	train	A drifting single qubit with two calibration strategies, trained off the main thread in a worker
Routing	optimise	SABRE and a REINFORCE-trained router on real coupling maps, with equivalence verification
Experiments	run	A lab notebook: hypothesis, steps with circuits, seeds and results, conclusion; persisted to Supabase or memory and also exposed as five MCP tools
Noise, Mitigation	measure	Trajectory noise models and readout-error mitigation by confusion-matrix inversion
QML Check	train	A variational quantum classifier versus logistic regression and a small MLP on three datasets, in a worker
GPU Simulator, Simulator Benchmark	run	CPU float64 versus WebGPU float32, capacity probing, and the correctness and performance suites for the engine
Run This Paper	run	A curated list of landmark results, each backed by an existing tested module, not an arXiv executor
Equivalence	measure	Checks two circuits for equivalence up to global phase
Device Advisor, Resource Estimator	measure	Device ranking from calibration data; order-of-magnitude fault-tolerant resource estimates
Entanglement Transition, Barren Plateaus	measure	Measurement-induced entanglement transition; gradient variance falling with qubit count (McClean et al. 2018)

Shared machinery

• verify.ts checks a circuit against a spec in the engine's own qubit convention, so a verification result means the same thing in the workbench, in the MCP generate_verified_circuit tool and in the challenge grader.
• exportBundle.ts builds a citable, re-runnable bundle (README.md, manifest.json, circuit.qasm, results.json, CITATION.bib). It is pure, takes the timestamp as a parameter and never reads the clock, so the same inputs always produce the same bytes, which is the property a reproducibility artefact must have. The circuit is emitted once, by toQasm, so there is exactly one OpenQASM emitter in the codebase.
• notebook.ts writes a Jupyter notebook (nbformat 4.5) targeting Qiskit or PennyLane with deterministic cell ids, so exports are byte-identical for the same circuit.
• circuitImage.ts renders SVG, PNG and quantikz LaTeX from the same geometry the editor uses.
• lib/numerics/nn.ts is a minimal shared neural-network toolkit (Float64Array parameters, Adam, Glorot initialisation) used by the decoders, zero-noise extrapolation and QML modules, so every learned component in the workbench trains with the same, tested optimiser.

Workers

Anything that trains runs in a Web Worker (routing.worker.ts, calibration.worker.ts, qml.worker.ts) with start and stop messages, periodic progress reports and budgets in episodes and seconds. The page stays responsive and a learner can abandon a run.

Why no arbitrary code execution

Several tools take text from a language model. None evaluates it. Generated circuits are parsed by fromQasm into a Circuit, evolved candidates are interpreted by a purpose-built DSL interpreter, and verification runs on the parsed structure. This is both a security boundary and what makes the tools testable: a scripted fake provider exercises every branch of the generate–verify–repair loop in milliseconds.