Skip to content

Mitigated Bell Re-Run Preregistration (KIMI-9)

Date: 2026-07-16

This preregistration prepares a re-run of the March 2026 Bell test with per-setting transparency and readout mitigation. 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, and executes only after the 2026-07-16 public-claim corrections are green on main.

Scientific question

The committed results/ibm_hardware_2026-03-28/bell_test_4q.json artefact carries an anomalous second analyser setting: E ≈ +0.286/+0.334 on the two pairs against ≈ 0.80–0.86 for the other three settings (confirmed by recomputation, scripts/recompute_chsh_bell_test.py; stated publicly in docs/results.md). Is that anomaly a readout/transpilation artefact, or a real asymmetry of the executed setting?

Claim boundary

Supported after successful execution and analysis:

  • per-setting correlators E with committed per-setting transpiled circuits;
  • CHSH S ± σ per pair, reported with readout mitigation on and off;
  • a dated live-surface note resolving or documenting the anomaly.

Blocked even after a positive result:

  • loophole-free or device-independent claims of any kind;
  • QKD viability claims (KIMI-8 tracks the QBER derivation separately);
  • any edit of the published March 2026 record — the original artefact and preprint stay as they are; this re-run is a NEW dated record.

Both outcomes are publishable: "anomaly was a correctable artefact" and "anomaly is real and the published S values stand with a documented setting asymmetry" have equal standing.

Circuit matrix

Field Value
Pairs 2 independent Bell pairs (4 qubits total), fresh layout from calibration
Settings 4 CHSH analyser settings, each a separate committed transpiled circuit
Main shots 4096 per setting
Readout calibration per-qubit readout-matrix circuits, 8192 shots
Mitigation readout-matrix correction on E and S; reported on/off
Analysis scripts/recompute_chsh_bell_test.py conventions (little-endian pairs, minus sign on setting 1, multinomial σ) applied to the new counts by a committed reproducer

Decision rules (preregistered)

  • If the mitigated setting-1 correlators rise to within 2σ of the other settings' band, the anomaly is explained as a readout/transpilation artefact and the live surfaces gain a dated resolution note.
  • Otherwise the asymmetry is documented as a real property of the executed setting, with the per-setting circuits published for inspection.
  • S is stated per pair with exact σ; no blanket significance statements.

Budget and abort criteria

  • Estimated total QPU wall: 2–3 minutes of the remaining ~18-minute budget (owner decision 2026-07-16: shared with WIDTH-1 and RC-1).
  • Abort before submission if fresh calibration readout error on any chosen qubit exceeds the mitigation circuit's correctable range.
  • Retrieval failing strict count coercion quarantines the run — partial evidence is recorded as partial.

Evidence protocol

Submission record, retrieval record, hash-bound result pack with raw counts and verifier, ledger row before public quotation, per the standing rules.

Status

  • [ ] NOW-block corrections green on main (prerequisite)
  • [ ] Owner GO with exact shot plan (per submission)
  • [ ] Executed / analysed / packed / ledgered