Repository navigation
Detector sampler raises IndexError when a non-Clifford block feeds no detector #260
Description
Hi! While fuzzing the statevector backend before 1.0.0, we found a small circuit that compile_detector_sampler can't sample. The non-Clifford part on qubits 0–1 isn't read by any detector or observable:
import tsim
c = tsim.Circuit("""
U3(0.25, -0.25, 0.825) 0
YCY 1 0
U3(0.125, 0.595, 0.5) 0
SQRT_YY 1 0
MR 1
X_ERROR(0.1) 2
M 2
DETECTOR rec[-1]
""")
c.compile_sampler(seed=0).sample(3) # works
c.compile_detector_sampler(seed=0).sample(5) # IndexError: index -1 is out of bounds for axis 0 with size 0Expected: 5 rows with one detector, which fires about 10% of the time.
What seems to happen: compile_program (src/tsim/compile/pipeline.py) still compiles the unread component, which then has rotations but no outputs and no symbols. lane_program (src/tsim/statevector/shared.py, lane_dormant = dormant & read[branch]) then indexes the empty read array with a branch of -1.
When the component does have symbols, read[-1] just reads an unrelated entry, and dormant masks it out. So the visible effect is the crash; we found no case of wrong samples.
Possible fixes: skip components without outputs when compiling for the detector sampler, or guard the index (e.g. read[np.maximum(branch, 0)] together with the existing dormant mask).
Random circuits hit this about once in 20k. It reproduces on main (e3a465b) and on #259 (a9af140).
🤖 Generated with Claude Code
Activity
Drop components without outputs before compiling
rafaelha commentedon Nov 4, 2026
A small correction and a few more triggers, from a 5-hour fuzzing run:
compile_samplercan crash too. It does when the circuit has no measurements at all. The fuzzer's smallest reproducer raises the sameIndexErrorfrom both samplers for any seed:CY rec[-2] 2) onto a non-Clifford qubit that is never measured again;SPP_DAGproducts on such qubits;HERALDED_ERASE,HERALDED_PAULI_CHANNEL_1, even at p=0) on a non-Clifford qubit that is never measured again. The fuzzer reported this one; I haven't reproduced it myself yet.In every case the cause is
read[branch]withbranch == -1inlane_program(src/tsim/statevector/shared.py:208). The fuzzer saw 18 crashes, 17 of them on this line. Across about 96k circuits checked against an exact oracle, it found no wrong samples.🤖 Generated with Claude Code
Drop components without outputs before compiling (QuEraComputing#262)