QBER Re-Run With Committed Basis Metadata (KIMI-8)¶
Date: 2026-07-17
This preregistration prepares a re-run of the March 2026 QBER measurement with per-pub basis metadata committed alongside the counts. It does not submit an IBM job, reserve backend time, or authorise QPU spend — submission additionally requires an owner GO with the exact shot plan.
Scientific question¶
The published QKD bit-error rate of 5.5% (ZZ) / 5.8% (XX) cannot be
re-derived from the committed
results/ibm_hardware_2026-03-28/qkd_qber_4q.json artefact because it
carries no basis metadata (dated caveat live on docs/results.md since
2026-07-16); a naive matched-basis sift of the committed counts gives
2.0–3.7%. What matched-basis error rate does the same entangled-pair
protocol give when every pub carries committed basis metadata, so the
derivation is reproducible from the pack alone?
Honesty boundary stated up front¶
This is a matched-basis mismatch measurement on entangled pairs (BBM92-style with fixed per-pub bases), NOT a QKD protocol run: there is no per-shot random basis choice, no sifting in the protocol sense, no authentication, and no privacy amplification. "QBER" here means the mismatch rate of the entangled source under matched measurement bases — a hardware-quality statement only.
Claim boundary¶
Supported after successful execution and analysis:
- matched-basis error rate per pair per basis (ZZ and XX), with binomial σ, readout mitigation on and off, and the exact basis metadata of every pub committed in the pack;
- a committed pure-arithmetic derivation of the March artefact's naive matched-basis sift (assumed pub bases stated explicitly) next to the new values;
- a dated live-surface note making the new record the citable error rate.
Blocked even after a positive result:
- any QKD security, key-rate, or viability claim;
- any claim that re-derives or repairs the published 5.5%/5.8% — the non-derivability caveat on the March record stands regardless of what the re-run measures;
- any edit of the published March 2026 record.
Circuit matrix¶
| Field | Value |
|---|---|
| Pairs | 2 independent Bell pairs (4 qubits), fresh layout from calibration (same selector as the KIMI-9 Bell re-run) |
| Main pubs | 2: matched ZZ (direct measurement) and matched XX (H on all four qubits before measurement), each a separate committed transpiled circuit with explicit alice_basis/bob_basis metadata |
| Main shots | 4096 per pub |
| Readout calibration | 16 full-basis circuits, 8192 shots (exact 2^n correction, mitigation/readout_matrix.py) |
| Analysis | pair marginal mismatch rate q = P(bits differ), binomial σ = sqrt(q(1−q)/N) at the raw shot count; mitigated rate from the exact readout inversion (quasi-probabilities enter unclipped, labelled) |
Decision rules (preregistered)¶
- The new matched-basis error rates are reported per pair per basis with exact σ, mitigation on and off. No blanket averages.
- If the new rates fall within the March naive-sift band (recomputed by the committed script from the committed counts) to within 2σ, the reading is: the published 5.5%/5.8% overstate what the March artefact supports, and the new dated record becomes the citable error rate.
- If the new rates are materially higher and near the published values, the reading is: the March analysis plausibly included effects the artefact does not record; the non-derivability caveat still stands, and the new dated record is still the citable error rate.
- Either way the caveat on the live surfaces stays; only its follow-up note changes.
Budget and abort criteria¶
- 18 circuits (2 main + 16 calibration) ≈ 20 estimated QPU seconds at the 1.1 s/circuit anchor; hard cap 60 s. Spend comes from the remaining ~1011 s of the approved budget and needs a per-submit owner GO.
- Abort before submission if fresh calibration readout error on any chosen qubit exceeds the abort threshold, if any transpiled depth exceeds the ceiling, or if the estimate exceeds the cap.
- Retrieval failing strict count coercion quarantines the run.
Evidence protocol¶
Submission record, retrieval record, hash-bound result pack with raw counts, basis metadata, and verifier, ledger row before public quotation, per the standing rules.
Status¶
- [x] Owner GO with exact shot plan — given 2026-07-17 (18 circuits, 4096/8192 shots, estimate 19.8 s, cap 60 s)
- [x] Executed / analysed / packed / ledgered — 2026-07-17, job
d9cosfcjeosc73fgikgg, 38 usage seconds; pack indata/qkd_qber_basis_metadata/, ledger row added