Error mitigation is not implemented yet
carcara.backends.mitigation is an empty stub. Shots submitted to a real QPU today run raw, with no zero-noise extrapolation or readout correction on top.
Hardware
Three backends, one number — and what still doesn't work
Every figure on this page is the project's own. The caveats below are kept as visible as the headline numbers on purpose: a result that hides its own limits is not more trustworthy for it.
Headline numbers
- max cross-backend spread
- 1.3×10⁻⁷ Ha max cross-backend spread Qiskit, Braket and Cirq vs. the internal NumPy state vector, LiH ADAPT-VQE
- measurement circuits
- 29 measurement circuits 118 Pauli terms of LiH, grouped by qubit-wise commutation
- QPU providers reachable
- 3 QPU providers reachable IonQ, IQM and Rigetti through the Amazon Braket device registry
- error mitigation, today
- 0 LOC error mitigation, today backends.mitigation is a stub — shots run raw on hardware
The same Hamiltonian, three SDKs
ADAPT-VQE on LiH — 8 fermionic excitation operators, 208 CNOTs — built from one shared gate stream (X, H, S, S†, CNOT, Rz) and translated per SDK. No exceptions, no cherry-picking: every backend against the same FCI reference.
| Backend | Ground-state energy | Error |
|---|---|---|
| NumPy (state vector) | −6.88824276 Ha | 1.34×10⁻⁷ Ha vs. FCI |
| Qiskit | −6.88824283 Ha | 6.27×10⁻⁸ Ha vs. FCI |
| Amazon Braket | −6.88824279 Ha | 1.09×10⁻⁷ Ha vs. FCI |
| Cirq | −6.88824281 Ha | 8.17×10⁻⁸ Ha vs. FCI |
"NumPy (state vector)" is the internal reference implementation, not an SDK — the row this table checks the other three against.
What doesn't work yet
Read these before the table above, not after.
Forces and geometry optimization are known-broken
Do not trust QuantumCalculator for structure relaxation yet — a water-molecule relaxation test diverges instead of converging. Single-point energies are unaffected.
Gradient screening on hardware is still classical
On a real QPU, only the final energy evaluation runs natively today; ADAPT-VQE's operator-gradient screening step still executes on the classical simulator behind it.
Checked independently
The caveats above are about what is missing. What exists is checked against results that have nothing to do with a single lucky run:
- Cross-backend state vectors agree to machine precision — the same anti-Hermitian generator factorizes exactly into Pauli rotations, no Trotter error
- H₂ and LiH ground-state energies checked against published FCI references
- A 609-test pytest suite covering integrals, mappings, drivers and backend-provider equivalence
- A dedicated Amazon Braket compatibility report — gate set, shots constraint, QWC grouping, QPU cost
Reaching real hardware
Real QPU access goes through Amazon Braket — an AWS account, IAM permissions for Braket, and backend_provider="braket", execute_circuits=True, device=<ARN>, shots=<n>. The install guide below covers the Python side; the how-to guide under docs/guide/aws_braket covers the AWS side.
What's next
A real error-mitigation backend, hardware-native gradient screening for ADAPT-VQE — today only the final energy evaluation reaches the QPU — and a GPU-accelerated integral backend alongside the existing C/OpenMP one.
Keep reading
Localized orbitals and pseudopotentials generated from scratch, and the C-accelerated integral engine underneath both.
VQE, ADAPT-VQE, excited states, VASQE and periodic systems — four solvers sharing one driver.