Skip to content

Full-Basis Readout Calibration Preregistration

Date: 2026-05-06

This preregistration prepares complete computational-basis readout calibration for DLA parity follow-up datasets where existing raw counts only support partial exact-state or parity readout correction. It does not submit IBM jobs, reserve backend time, or authorise QPU spend.

Scientific Question

For promoted or follow-up DLA parity datasets, does a full computational-basis readout-confusion correction preserve the sign and magnitude of the leakage asymmetry beyond the partial exact-state corrections already implemented?

Existing Eligibility Boundary

The current eligibility markers are generated by:

python scripts/audit_readout_mitigation_eligibility.py

Current output:

data/readout_mitigation_eligibility/readout_mitigation_eligibility_2026-05-06.json

The markers classify committed n<=8 raw-count datasets as:

  • full_basis_eligible;
  • partial_exact_state_baseline_only;
  • missing_readout_calibration.

New calibration circuits are justified only for datasets or planned follow-up runs where the scientific decision depends on full confusion-matrix inversion.

Claim Boundary

Supported after successful execution and analysis:

  • full 2^n x 2^n readout-confusion matrix construction for the calibrated qubit subset;
  • comparison of raw, partial exact-state corrected, and full-basis corrected leakage metrics;
  • statement that a promoted sign either survives or fails under full-basis measurement mitigation.

Blocked even after calibration:

  • mitigation of gate errors, Trotter error, decoherence, or crosstalk;
  • quantum advantage;
  • backend-general readout behaviour;
  • reuse of a calibration matrix across a different backend, qubit layout, or calibration window without explicit drift justification.

Calibration Matrix

For each selected n and physical layout:

n Basis states Shots per state Calibration shots
4 16 8192 131072
6 64 4096 262144
8 256 2048 524288

Default approval target: n=4 only, because it is cheap and directly supports the promoted DLA parity and state/layout controls.

Escalate to n=6 or n=8 only if:

  • the target paper claim explicitly depends on that size;
  • the backend/layout is fixed;
  • the expected QPU time remains justified by the decision value;
  • the user explicitly approves the calibration spend.

QPU-Time Estimate

Expected IBM-reported QPU time depends on backend scheduling overhead and readout-circuit duration. Conservative block ceilings:

  • n=4: 1-3 QPU minutes, ceiling 5 minutes;
  • n=6: 3-8 QPU minutes, ceiling 10 minutes;
  • n=8: 10-25 QPU minutes, ceiling 30 minutes.

Abort before submission if the live estimate exceeds the ceiling or if the remaining allocation cannot cover the block plus a 25 % safety margin.

Live Readiness Gates

Before submission:

  • confirm target dataset or planned follow-up run needs full-basis mitigation;
  • confirm backend, physical qubits, and layout are fixed;
  • generate all basis-state preparation circuits from committed code only;
  • live-transpile all calibration circuits;
  • reject if any calibration circuit maps outside the target qubit subset;
  • record backend name, physical qubits, calibration timestamp, circuit count, shot count, estimated QPU minutes, and transpiled depth summary;
  • get explicit approval immediately before submission.

Analysis Plan

For each calibrated layout:

  • build the empirical assignment matrix;
  • report row retention, parity-flip rate, and condition number;
  • compare raw and mitigated leakage metrics;
  • flag unstable inversions or high condition numbers rather than forcing a correction;
  • store the matrix and correction outputs as committed JSON/CSV artefacts.

Use full-basis correction only when the matrix is numerically stable and the calibration window matches the analysed raw-count dataset closely enough to be scientifically defensible.

Falsification Rules

The existing leakage claim must be downgraded if:

  • full-basis correction reverses the promoted sign;
  • the effect collapses inside calibration uncertainty;
  • the assignment matrix is too ill-conditioned for a stable inversion;
  • calibration drift makes the correction inapplicable.

If full-basis mitigation is inconclusive, retain raw and partial-correction claims only with the existing limitation language.

Output Artefacts

Expected paths after approved execution:

  • data/readout_full_basis/readout_full_basis_<backend>_<n>q_<timestamp>.json;
  • data/readout_full_basis/readout_full_basis_matrix_<date>.csv;
  • data/readout_full_basis/readout_full_basis_summary_<date>.json;
  • docs/campaigns/readout_full_basis_manifest_<date>.md.

Each artefact must include backend, physical qubits, layout, job ID, raw counts, assignment matrix, SHA256 hashes, condition number, and reproduction commands.

Submission Boundary

This preregistration is complete. QPU calibration execution remains blocked until a specific dataset/layout requires full-basis mitigation, live readiness artefacts are generated, budget is confirmed, and explicit approval is given.