For researchers

You care about what a device actually does, and about being able to say exactly what you ran. The lab is built to make both checkable.

#Reproducibility, by construction

• A project is files, and every run is submitted from a .qasm you can read. The Runs rail keeps shots, machine and cost per run; the run detail keeps the submitted circuit and the counts. Nothing is priced or transformed client-side.
• The codename you see is the machine's public targetId — the same string in every SDK, so a run description in a paper ("<codename>, 4096 shots") is enough for someone else to repeat it.
• The circuit builder refuses to draw anything it cannot draw honestly (custom gates, multiple registers, QASM 3 keep you in code with the parser's own reason). The emulator is an ideal statevector — no noise, no sampling — and says so on its header; it stops at 16 qubits in the browser rather than pretend, and past that it points you at the server simulator that fits (22 qubits per the catalogue) instead of a dead end.
• The library quotes observed output for every runnable item and marks reference material Reading only with licence and upstream path.

#Simulator vs device, side by side

The GHZ template is the canonical experiment: run ghz.qasm on a simulator and on a real machine at the same shot count, then open both runs from the Runs rail. The simulator gives 000 and 111 and nothing else; the device gives the other six outcomes too, and their share is a rough, honest read of what three entangling gates cost on that machine today. The histogram sorts outcomes and expands to all of them; the credit line says what was held and settled.

#Python where you need it

The browser kernel is CPython 3.12 with numpy, matplotlib and the real roro_platform SDK — including LocalSimulator, noise-model and mitigation plugins, Observable expectation values and SPSA in roro_platform.optimize — so parameter sweeps and small VQE/QAOA loops run offline in a notebook, free, with figures under the cells. Compiled stacks (qiskit, scipy, pennylane) are not available in the browser; write against roro_platform here and run the same code on your machine with the platform SDK from the RoRo repository (not on PyPI yet). Every result you'd cite is exportable: Runs rail → ⋯ on a run → Manifest (JSON) (machine, shots, QASM + SHA-256, counts, cost) or Counts (CSV/JSON) / Histogram (SVG); pick 2–4 runs with Compare for side-by-side histograms and pairwise total-variation distances; library items carry a Cite section (BibTeX + plain).

#Keeping the record

• Share produces an immutable snapshot link — a citation-grade pointer to exactly the code you ran, revocable if needed.
• Notebooks save outputs into the file, so a shared notebook carries its plots.
• Ask Explain this result for a first-pass narration; treat it as a draft. Copilot's transcript shows every file it read and every diff you approved, and its cost per turn.

#Limits to know

In-browser simulation caps at 16 qubits — both the kernel's LocalSimulator and the builder's emulator (the server simulators take 22); the kernel has no threads or subprocesses; a .py cannot submit to hardware from inside the lab — that goes through the .qasm Run rail, which quotes and confirms first.

Related: Circuit builder & emulator · Notebooks · Machines & credits