Skip to content

IBM v2 raw-count recovery and full disclosure (2026-07-18)

Summary

The "IBM v2" fair-experiment campaign (2026-03-29, ibm_fez) was originally committed aggregate-only — fidelity/survival means plus HMAC-blinded run labels, with no raw counts and no raw job identifiers. Because it could not be independently checked, it was quarantined (see Count-Integrity Incident, April 2026).

On 2026-07-18 the nine jobs were re-retrieved read-only from IBM (0 QPU seconds) and published in full. This document records the recovery method, the reproduction result, and the claim boundary.

What is now public

data/ibm_hardware_v2_recovered_2026-07-18/:

  • recovered_raw_counts.json — for each of the nine experiments: the real IBM job identifier, the per-pub raw counts (10 pubs × 8192 shots = 81 920 shots each, 4 measured qubits), and the dated ibm_fez calibration snapshot (last_update 2026-03-29).
  • manifest.json — SHA-256 of the pack.
  • reproduction_analysis.json — the reproducer output.

Full-disclosure policy (owner directive, 2026-07-18): raw counts, real job identifiers, and calibration are public research data. The only value withheld is the IBM API token; the recovery tool runs a fail-closed leak check against it before writing anything.

Method (reproducible)

  1. scripts/recover_ibm_v2_raw_counts.py authenticates read-only, enumerates the ibm_fez job cluster submitted 2026-03-29 ~12:20 UTC (nine jobs in ~10 s), pulls per-pub raw counts and the dated calibration, pins each job to its experiment by submission order cross-checked against the committed aggregate, and writes the pack.
  2. scripts/analyse_ibm_v2_recovered.py recomputes each experiment's survival observable from the raw counts (all-zero survival for the ground-state preparations A_even, C_xy, C_fim, B_M+4; the sector target state for B_M±; the odd-parity subspace for A_odd) and compares to the committed value.

Reproduction result

Eight of nine experiments reproduce to |Δmean| < 1×10⁻⁴ directly from the raw counts. A_odd reproduces to ~3.7 % under the natural odd-parity-subspace definition — its committed aggregate applied an original readout mitigation whose exact form is not in the pack; the residual is reported, not fitted away.

Experiment Committed mean Recovered mean Observable
A_even 0.9185 0.9185 P(0000)
A_odd 0.8526 0.8900 odd-parity subspace (±3.7 %)
C_xy 0.8484 0.8484 P(0000)
C_fim 0.9158 0.9158 P(0000)
B_M+4 0.9936 0.9936 P(0000)
B_M+2 / M0 / M−2 0.0000 0.0000 sector target
B_M−4 0.9595 0.9595 P(1111)

The committed aggregates are therefore genuine, not fabricated — the April 2026 count-integrity incident concerned a separate large-N frontier automation lane, not these nine jobs.

Claim boundary

Reproducing F_FIM = 0.9158 > F_XY = 0.8484 from real hardware counts establishes only that the observation is genuine. It is not evidence of a coherence-protection ("DUAL PROTECTION") mechanism: the FIM and XY circuits differ in depth and structure, so a higher all-zero survival for the FIM circuit is consistent with a shallower/less-decohering circuit rather than any protection effect. The coherence-protection hypothesis was tested properly on the promoted ibm_kingston SCPN/FIM campaign and committed as a negative/falsification result (digital λ=4 increases leakage / decreases retention; claim boundary). The "DUAL PROTECTION on IBM hardware" wording is therefore retired as a claim in data/retired_claims.json (fim-dual-protection-hardware) while the data stands.