Skip to content

Next validation protocols -- no-submit preparation

Date: 2026-05-05

This document prepares the next validation steps without submitting new IBM jobs. The objective is to spend zero additional QPU time until offline checks justify which hardware run is worth the remaining allocation.

1. Readout-mitigation cross-check

Status: implemented as scripts/analyse_phase2_readout_mitigation.py.

The existing Phase 2 data do not contain a complete computational-basis confusion matrix. Therefore a literal full 2^n x 2^n confusion-matrix inversion is not available from current raw counts. The implemented cross-check performs state-specific parity-confusion inversion for rows whose initial state has an exact readout-only calibration.

Decision rule:

  • If sign and Fisher significance survive parity-readout correction, keep the manuscript claim unchanged and add a short robustness sentence.
  • If correction changes the sign of a promoted claim, downgrade the claim before any submission.
  • A true full confusion-matrix inversion requires a new calibration block with all basis states for the measured qubit subset.

2. Offline GUESS / symmetry-decay calibration

Status: implemented as scripts/analyse_phase2_guess_calibration.py.

The current datasets do not contain explicit noise-scale folding, so the analysis is not a GUESS zero-noise extrapolation result. It fits parity survival versus circuit depth only as a readiness check for whether parity leakage is smooth enough to serve as a future symmetry-decay witness.

Decision rule:

  • If several state/sector series have high log-linear fit quality, prepare a minimal folded-noise GUESS validation run.
  • If fit quality is mixed, keep GUESS as future work only.

3. State/layout-randomization QPU protocol

Goal: separate symmetry sector, excitation count, and physical layout effects.

Prepared design:

  • Use n=4 only.
  • Depths: d in {6, 8, 10, 14}.
  • States: 0011, 0101, 0001, 0010, 0111.
  • Layouts: three connected four-qubit windows selected by live backend coupling map and lowest recent readout error.
  • Repetitions: 8 per state/depth/layout.
  • Shots: 4096.
  • Readout calibration: all five prepared states per layout, 8192 shots.

Estimated size:

  • Main circuits: 5 states x 4 depths x 3 layouts x 8 reps = 480.
  • Readout circuits: 5 states x 3 layouts = 15.
  • Total: 495 circuits.

Estimated QPU time:

  • Based on the completed popcount-control run, this is expected to cost roughly 8--15 QPU minutes, depending on selected backend queue and transpiled depths.

Submission rule:

  • Do not submit until live transpilation confirms max depth below the existing popcount-control depth envelope and the user explicitly approves QPU spend.

4. Minimal multi-device replication protocol

Goal: determine whether the n=4 asymmetry is specific to ibm_kingston or stable across another Heron backend/calibration.

Prepared design:

  • Backend: second Heron r2 backend only.
  • Scope: reduced A block, no scaling.
  • Depths: d in {4, 6, 8, 10, 14, 20}.
  • States: original 0011 even and 0001 odd.
  • Repetitions: 12 per state/depth.
  • Shots: 4096.
  • Readout calibration: 0011, 0001, plus 0000, 1111, 8192 shots.

Estimated size:

  • Main circuits: 2 states x 6 depths x 12 reps = 144.
  • Readout circuits: 4.
  • Total: 148 circuits.

Estimated QPU time:

  • Using observed Heron rates from Phase 2, expected cost is roughly 3--6 QPU minutes if live transpilation depths remain comparable.

Submission rule:

  • This is the preferred next hardware run only if the publication strategy requires backend-transfer evidence before submission. Otherwise keep it as a post-preprint replication.