Phase 2 no-QPU cross-checks¶
Date: 2026-05-05
These checks spend no additional IBM QPU time. They use only promoted Phase 2 raw counts and derived summaries.
Readout mitigation boundary¶
Script: scripts/analyse_phase2_readout_mitigation.py
Output:
data/phase2_readout_mitigation/phase2_readout_mitigation_summary_2026-05-05.json
The existing raw counts include selected readout-only calibration states, not a
complete computational-basis calibration set. Therefore the repository cannot
honestly claim a full 2^n x 2^n confusion-matrix inversion without new QPU
calibration circuits.
Implemented correction:
- exact-state parity-confusion inversion,
- only for initial states with matching readout-only calibration rows,
- no full-count redistribution across all basis states.
Main result after correction:
- Phase 2 A+G
n=4replication: Fisherchi2=73.894393,p=4.160825e-08, significant depths5/10. - Popcount original contrast: Fisher
p=9.267543e-18. - Popcount within-even swap: Fisher
p=4.466548e-20. - Popcount within-odd swap: Fisher
p=3.590153e-15. - Popcount excitation-inversion arm: Fisher
p=5.929924e-18.
Conclusion: readout-only parity correction weakens the Phase 2 A+G combined statistic but does not remove the promoted sign/statistical conclusion. The popcount-control downgrading of the causal claim also survives correction.
Offline GUESS-readiness calibration¶
Script: scripts/analyse_phase2_guess_calibration.py
Output:
data/phase2_guess_calibration/phase2_guess_calibration_summary_2026-05-05.json
This is not GUESS zero-noise extrapolation because the promoted Phase 2 datasets do not include explicit folded-noise scale factors. The analysis fits log parity-survival versus circuit depth as a cumulative-noise proxy.
Fit quality:
- Phase 2 A+G even:
alpha=0.034190,R2=0.9989. - Phase 2 A+G odd:
alpha=0.034662,R2=0.9997. - Phase 2 A+G sector mean:
alpha=0.034427,R2=0.9995. - Popcount state fits:
R2=0.9582to0.9921.
Conclusion: parity leakage is smooth enough versus depth to justify preparing a future GUESS-style folded-noise protocol. It is not yet a demonstrated error mitigation result.
Hardware follow-up preparation¶
Prepared protocol:
docs/campaigns/next_validation_protocols_2026-05-05.md
Recommended order:
- Keep the current paper as a no-new-QPU preprint after incorporating the readout and GUESS-readiness boundaries.
- Use the state/layout-randomization protocol only if reviewers or collaborators ask for mechanism separation beyond the popcount control.
- Use the minimal multi-device protocol only if backend-transfer evidence is required before journal submission.