Translation validation for the Qiskit transpiler: verify that a
PassManager's output implements the same operation as its input,
pass by pass, on the Clifford fragment.
A transpiler pass that silently produces a non-equivalent circuit is a correctness bug that's easy to miss in review and hard to catch with unit tests alone, since the bug only shows up on specific gate sequences. This tool checks every pass in a run against stabilizer tableau equivalence, so a miscompilation is caught at the pass that introduced it — not discovered downstream as a wrong measurement distribution.
qiskit-transpile-verify v0.1 verifies circuits made up of Clifford gates
(h, s, sdg, sx, x, y, z, cx, cy, cz, swap, ecr, and
rz/rx/ry/p at Clifford angles). It does not verify circuits
containing t, arbitrary rotation angles, or other non-Clifford gates —
those are reported as out-of-scope, not silently skipped or falsely
passed. See docs/LIMITATIONS.md for exactly what
this does and doesn't prove, and why universal (Clifford+T) verification
is a substantially different — and harder — problem, not just a bigger
gate list.
pip install qiskit-transpile-verify
# CLI (QASM3 file support) needs one extra dependency:
pip install "qiskit-transpile-verify[cli]"from qiskit.circuit import QuantumCircuit
from qiskit.transpiler import PassManager
from qiskit.transpiler.passes import CommutativeCancellation
from qiskit_transpile_verify import verify_transpile
qc = QuantumCircuit(3)
qc.h(0)
qc.cx(0, 1)
qc.cx(0, 1) # redundant pair
qc.cx(1, 2)
qc.s(2)
report = verify_transpile(PassManager([CommutativeCancellation()]), qc)
print(report.summary())
assert report.all_verified1 pass(es) run:
[0] CommutativeCancellation: VERIFIED
This isn't a hypothetical. Here's a deliberately broken cancellation pass
that cancels cx(a, b) against cx(b, a) as if they were inverses of
each other — they aren't, since swapping control and target changes the
operation — run through verify_transpile:
class BuggyCXCancellation(TransformationPass):
"""Cancels any two adjacent CX gates on the same qubit pair,
regardless of control/target order. This is wrong."""
...
qc = QuantumCircuit(2)
qc.h(0)
qc.cx(0, 1)
qc.cx(1, 0) # NOT the inverse of cx(0, 1)
report = verify_transpile(PassManager([BuggyCXCancellation()]), qc)
print(report.summary())1 pass(es) run:
[0] BuggyCXCancellation: MISMATCH — possible miscompilation
FIRST FAILURE at pass [0] BuggyCXCancellation: no qubit permutation
reproduces a matching tableau
The full runnable version of this example is in tests/test_verify.py::test_buggy_pass_is_caught.
The CLI runs a QASM3 circuit through one of Qiskit's built-in preset pass managers and verifies every pass:
qtv circuit.qasm3 --optimization-level 337 pass(es) run:
[0] ContainsInstruction: VERIFIED
[1] UnitarySynthesis: VERIFIED
[2] HighLevelSynthesis: VERIFIED
...
[36] ContainsInstruction: VERIFIED
(Tested against Qiskit 2.5.2's generate_preset_pass_manager at
optimization levels 0–3 on Clifford-fragment input; every individual pass
verified across all levels tried so far. This is not a formal proof that
the preset pass managers are bug-free in general — see Limitations.)
verify_transpile(pass_manager, circuit)— run and verify a wholePassManager, pass by pass. Returns aVerificationReportwith.all_verified,.first_failure,.out_of_scope_passes,.summary().check_equivalence(circuit_a, circuit_b, allow_permutation=True)— check two circuits directly, with optional qubit-permutation search (useful since layout/routing passes legitimately relabel qubits).check_equivalence_with_layout(circuit_a, circuit_b, layout)— same, but with a known layout instead of brute-force search (use this above ~8 qubits, since permutation search is O(n!)).check_scope(circuit)— check whether a circuit is in the Clifford fragment before attempting verification.
This package extends the translation-validation idea from
qiskit-zx-verified —
which formally proves (Lean 4, zero sorrys) that one specific Clifford
phase-fusion pass is correct — into a general-purpose checker that works
on any PassManager, at the cost of a weaker guarantee: a fast runtime
equivalence check rather than an independently checkable proof term.
Generating Lean certificates for arbitrary passes is the planned v0.2
direction; see lean/README.md for the honest current
status (not implemented in v0.1) and the reasoning behind the scoping.
Apache-2.0