Tensorfix

Given paired reference and non-equivalent output circuits, it extracts a Semantic Failure Witness (SFW), enumerates bounded local edits, ranks them by process fidelity and structural cost, and accepts a patch only when MQT QCEC and a strict small-circuit operator cross-check agree.</p><h2 dir="ltr">Repository map</h2><ul><li>src/witness: SFW extraction, layouts, and interaction closure.</li><li>src/tensor: dense and Choi tensor-network process fidelity.</li><li>src/repair: candidate generation, ranking, local search, and paired full-circuit localization/splicing.</li><li>src/validation: fail-closed QCEC plus strict small-circuit operator checking.</li><li>src/baselines: rollback, four shared-pool search orders, and independent Qiskit resynthesis.</li><li>experiments: main, ablation, full-circuit integration, multi-seed, multi-gate stress, provenance-backed regressions, and measured scalability.</li><li>results/raw: generated experimental records.</li><li>analysis/process_results.py: summary, macro, and figure generation.</li></ul><h2 dir="ltr">Reproduce</h2><p dir="ltr">Install the pinned Python environment:</p><pre><pre>pip install -r requirements.txt<br></pre></pre><p dir="ltr">Validate the dataset and run the audited experiments:</p><pre><pre>python scripts/validate_dataset.py<br>python scripts/run_full.py<br></pre></pre><p dir="ltr">Reproduce the networked multi-version Qiskit study (cached virtual environment per release):</p><pre><pre>python scripts/run_historical.py<br></pre></pre><p dir="ltr">Regenerate processed results and manuscript numbers:</p><pre><pre>python -m analysis.process_results<br></pre></pre><p></p>