Backends

Backend specs are strings so test files stay declarative and CI-safe.

Spec	What runs	Needs
aer (default)	Ideal (noiseless) local simulation via Qiskit Aer	nothing
ibm:<device>	Local Aer simulation under the named IBM device's bundled calibration snapshot (real historical noise data: gate errors, readout errors, T1/T2)	pip install 'qontinuum[ibm]'
ibm:<device>@live	Same, but with the device's current calibration fetched at run time	a saved (free) IBM Quantum account

@qtest(shots=4000)                          # ideal
def bell(): ...

@qtest(shots=4000, backend="ibm:manila")    # offline, real noise profile
def bell_noisy(): ...

@qtest(shots=4000, backend="ibm:brisbane@live")  # today's calibration
def bell_today(): ...

Device names are matched leniently: manila, fake_manila, and ibm_manila all resolve to the same snapshot. Unknown names fail with the list of available devices.

Setting thresholds under noise

Ideal-simulator thresholds (e.g. tvd_threshold=0.05) are usually too strict under a real noise model. A practical workflow:

1. Run once with qont test against the noisy backend and look at the reported TVD.
2. Set the threshold with headroom above the observed value but below "broken" (a missing gate typically moves TVD by 0.3+, an order of magnitude above noise).
3. Or use snapshot=True and let the two-sample test track the noisy distribution itself — then calibration-snapshot updates are reviewed like any other diff.

Why simulate noise instead of running on hardware?

CI needs to be fast, free, and deterministic-ish. Real-hardware execution (queues, cost, calibration drift) belongs in scheduled/gated jobs, not on every push — that's what qont run --on <provider:device> --max-cost <usd> is for: the whole suite is priced up front and nothing is submitted unless the estimate clears the budget (default $0). Noise-model simulation from real calibration data catches the class of bugs unit tests can: logic regressions, and "this circuit is now too deep to survive realistic noise."