qtwist 7.0.0

Certified-methodology quantum emulation SDK: every number ships with its typed error bound and a machine-checkable certificate — or the request is refused in milliseconds with a reasoned argument. Supersedes the 5.x legacy line (clean major break, same import name).

Quickstart

pip install qtwist==7.0.0

from qtwist import QTWIST, Circuit, Observable

sdk = QTWIST()                       # mock Heron twin (default)
c = Circuit(3)
c.h(0); c.cx(0, 1); c.cx(1, 2)       # 3-qubit GHZ

out = sdk.query(c, Observable.zz(0, 2), eps_target=0.05)
print(out.result)                    # value +- typed bound [CERTIFIED]
print(sdk.verify(out.result.certificate, program=c))   # True

Expected: value +0.910750, bound 0.024500 [CERTIFIED], and standalone certificate verification returning True.

How to read a result (two independent axes)

value       : +0.910750
bound       : 0.024500 [CERTIFIED]   <- eps-grade axis: the interval
                                      value +- bound is certified by
                                      typed bound arithmetic, GIVEN
                                      the twin's noise model
status      : Uncertified [Mock Calibration]   <- twin-trust axis:
                                      the default backend is a mock,
                                      so nothing is claimed about any
                                      real device

Tight bound on a mock twin is honest about being a mock; a loose bound on a measured twin is honest about model uncertainty. Neither axis borrows credibility from the other.

Core calls

Rejections are results, not exceptions: ask for a tolerance below the calibration floor and you get INFEASIBLE_TARGET in ~7 ms with the full term decomposition (eps_model, eps_drift, eps_readout_xtalk, ...) and actionable suggestions. Q-TWIST never returns a wrong value in place of a failure.

Registering a measured calibration

The default twin is a mock. To run against your own measured calibration:

from qtwist import (QTWIST, CalibrationLake, CalibrationRecord)

lake = CalibrationLake()
lake.add(CalibrationRecord(key="readout.e0e1",
                           values=(0.008, 0.016), ci=(0.002, 0.002),
                           source="experiment:my-lab",
                           method="confusion-matrix",
                           timestamp="2026-09-05 00:00:00+00:00"))
snap = lake.snapshot(device="my-device-v1",   # non-fake name
                     drift_per_hour=0.001, staleness_hours=1.5)
sdk = QTWIST(snapshot=snap)      # status now: Certified [OK]

Every record carries values + CIs + source + method + timestamp: provenance is part of the type system. A live SPRT drift monitor (sdk.observe(...), sdk.twin_status) rides on top.

What's inside

• 8 engines behind one certified oracle: quantum trajectories (A), exact density-matrix (B), Pauli-DAOE (C), adaptive MPS (D), TDVP accuracy-first (H), CHIMERA geometry synthesis (X), analog/DAQC (Q), logical FTQC twin (L).
• Twelve-term typed epsilon budget; default confidence 1e-3 delta; composition theorem for stacked queries.
• Signed Emulation Certificates (qtwist.cert.v1), standalone verifiable without re-running the simulation.
• Circuit API: h x y z s sdg t rx ry rz sx u3 cx cz swap iswap rxx ryy rzz rzx ccx; observables z, zz, Pauli sums.
• Requires Python >= 3.10 and numpy. Nothing else.

Status (honest)

This is an existence release: the SDK is parked here to hold the name and the version line. It is complete and audited (adversarial audit 73/73; hardening lineage 80/80; 43-check robustness battery; 23/23 probe suite; 9,480-line certified core, hash-pinned; release commit 77b562c, v7.0.0), but it ships mock device calibration only (fake-heron) — no real-QPU backends in the wheel. Real-device validation evidence lives in the external ARENA ledger: two IBM-marrakesh campaigns are on record (layout-pinned; campaign #0002 coverage 6/6 by interval overlap at ±0.05, 4/6 under the stricter point-containment reading with two near-misses — erratum E1 on record); the current arena certificate is PROVISIONAL pending the drift-poll and depth-ladder upgrade now in progress. Certificates use demo-grade HMAC signatures; production Ed25519 signing is the 7.1 roadmap.

If you found this package before the announcement: it is real, tested, and safe to try — it is simply not the launch.

Lineage

5.x (2026-08) was the pre-hardening product line with a web console; it remains on PyPI as the legacy museum piece. 7.0.0 is the certified-instrument line: different contract, numpy-only, one public name.