# Referee report: three-box "negative weak value ~ -0.53" (ibm_fez, 11 Oct 2026) Scripts and outputs are in this directory: load.py, a2.py (fit), noisy.py (+noisy_res.json), param.py/fitp.py, dilute.py, cmp.py, dup.py, off.py, layout.py. ## Verdicts - (a) NARROW. Wording: "On ibm_fez the corner box reads /(2 eps) = -0.48 +-0.02 (eps 0.4), -0.54 +-0.03 (0.3), -0.59 +-0.04 (0.25), -0.55 +-0.07 (0.2), -0.49 +-0.09 (0.15), -0.60 +-0.14 (0.1). The same transpiled circuits, noiseless, give -0.80 to -0.97 over this range, so the ratio to the circuit's own ideal is about 0.6, not about 0.5 of textbook -1. No eps dependence is resolved (WV = -0.66 + 0.43 eps, slope +-0.20)." "Flat" is a non-detection, not a measurement. -0.53 is a fit coefficient of an ad hoc form (see below), not a weak value. - (b) DROP. A common additive probe offset (+0.05, measured in the unused box where the ideal is exactly 0) plus FakeFez noise reproduces all 24 (box, eps) points, corner included: chi2 8.7 for 6 points (41.6/22 over all). Raw S+P fit of FakeFez is P=-0.674+-0.013 against -0.530+-0.053 (2.7 sigma), but that gap is the offset that the FakeFez snapshot lacks. Noise models differ by chip and layout: corner at eps 0.25 spans -0.68 to -0.81 (FakeMarrakesh -0.755..-0.808, FakeFez -0.68..-0.73, FakeKingston -0.73..-0.77 over 6 transpiler seeds); hardware -0.51 (marrakesh), -0.50 and -0.59 (fez). - (c) DROP as "specific / unexplained"; keep as a ~2 sigma observation. Ratio of fitted corner/positive coefficients: hardware 0.65 +-0.08; ideal circuit 0.87; FakeFez 0.84. With ideal finite-strength response divided out, per-box gains are A 0.846, B 0.777, corner 0.723 (+-0.025 each); mean(A,B) minus corner = 0.089 +-0.039. Including the measured excess post-selection acceptance (hardware corner keep 0.157 vs ideal 0.125 at eps 0.25) the gap is 0.03 +-0.05. FakeFez alone already reduces the corner more than the positives (gain 0.73 vs 0.85). On 24 Sep marrakesh (standard variant) the corner (-0.689) is larger in magnitude than the positives (0.579, 0.579). - (d) NARROW. Wording: "A positive additive offset of the probe, +0.062 +-0.010 (post-selected, unused box, constant fit chi2/dof 0.74, 6 sigma), +0.041 +-0.004 over all shots, is present in every box; the four null-box jobs of 24 Sep give +0.060 +-0.016. FakeFez gives +0.007." S=+0.111 is not a stable number: it comes from a free-P fit in which P=-0.10+-0.06 is correlated with S; the offset is common to all boxes (corner, A, B give S -0.02, +0.07, +0.03 in the same fit), not specific to unused. Consistent with a probe read-out asymmetry b-a of 0.04 to 0.06; this was not tested directly. ## 1. Recompute - All 9 npz recomputed. bit0=c0 (q142), bit1=c1 (q143), bit2=c2 (q144) agrees with the qasm lines `c[0]=measure $142; c[1]=measure $143; c[2]=measure $144`. Control: swapping probe to bit0 gives corner keep 0.07 and positive Xp; the ideal corner keep is 0.146 and hardware gives 0.173 with bit2 as probe. Post-selection c0=c1=0 matches the circuit (CZ then prep-dagger sends phi to |00>; ideal keep = 1/9 for A, B, unused, and 0.146 (eps 0.4) down to 0.113 (eps 0.1) for the corner). - eps from qasm: rz($144) = pi - eps/2 gives 0.4, 0.3, 0.25, 0.2, 0.15, 0.1 for circuit sets 62, 90, 118, 146, 174, 202. Structure is identical across sets (diff after normalising the angle: identical for all 4 boxes in sets 62, 90, 118; box A identical in 146, 174, 202). - GAPS IN THE RECORD: qasm exists only for jobs ti2, ti6, tia (all 4 boxes) and for box A of tjs, tkf, tkn. The three duplicate jobs (tjf, tjj, tjn) and boxes B, corner, unused of sets 146, 174, 202 have no qasm. eps for them is assigned from the circuit names in oct11_unrecorded_detail.json. Qubits 142-144 are known only from the qasm that exists. - Claimed table reproduced to the digit: A S=+0.073+-0.029 P=+0.817+-0.054 chi2/dof 2.28; B +0.033/+0.819/0.92; corner -0.018/-0.530/0.81; unused +0.111/-0.101/0.43. ## 2. Simulation (the qasm as transpiled; 3 physical qubits; Aer; 200k shots, FakeFez noise) Xp at eps 0.4..0.1 and S+P fit through the same form: - Ideal: corner Xp = -0.642, -0.528, -0.450, -0.371, -0.286, -0.201, fit S=-0.038 P=-0.852. Positive boxes: Xp = sin(2 eps) exactly (P about 0.98). The corner is NOT of the form S+P sin(2 eps) even without noise. WV at 0.25: A +0.959, corner -0.900. - FakeFez full: A S+0.028 P+0.802; B +0.002/+0.858; corner -0.010/-0.674; unused +0.007/+0.016. WV at 0.25: +0.808, +0.824, -0.658, +0.050. - Components at eps 0.25 (A / corner WV): readout only +0.915/-0.867; CZ+1q gate error only +0.837/-0.739; thermal only +0.953/-0.829. - So ordinary noise reproduces the positive boxes (0.80 to 0.86 vs hardware 0.82) and takes the corner to -0.67, hardware -0.53 (offset difference, see (b)). ## 3. Ordinary causes - Same CZ count (16), sx count (36) in all four boxes; X count 2 to 4. Gate error and T1 both lower the corner proportionally more than the positives (corner/positive WV ratio at 0.25: ideal 0.94, readout 0.95, gates 0.88, thermal 0.87, all 0.81). Corner has a larger post-selection acceptance (the probe is entangled with the system in the strong-coupling regime), and noise-induced false acceptances (hardware 0.157 vs ideal 0.125) dilute it. - Parametric 3-qubit model (gate scale, T scale, per-qubit read-out asymmetry) fit to A, B, unused (Xp and keep) predicts corner Xp within z=-2.2 to +1.5 at all six eps (sum z^2 9.1/6), but does not fit the acceptance fractions (corner keep chi2 214/6): the model is incomplete, and the keep excess is an unexplained observable. - A single global gain (0.775) plus offset (0.073) on the ideal circuit fits all 24 points: chi2 37.2/22. ## 4. Fit form, pooling, offset - Duplicate jobs (62, 90, 118, all 4 boxes, 12 pairs): chi2 = 11.1/12 (p=0.52). Pooling is valid. Per-job fit (9 points per box) gives the same numbers: A 1.60/dof, B 1.03, corner 1.05, unused 0.45. A pooled chi2/dof 2.28 (p about 0.06) is driven by eps 0.15 (z=-2.9 against the form). - Form: Xp = sin(2 eps) is exact only for positive boxes. For the corner it is wrong at the noiseless level (S -0.04, P -0.85) and the S parameter absorbs both this and the probe offset. Constant-only (no eps dependence in Xp) fails (chi2/dof 13 to 30), so Xp does scale with eps. S=0, P-only: A 0.946 (chi2/dof 2.2), B 0.877 (1.04), corner -0.562 (0.97), unused +0.104 (2.0, rejected). - Slope of WV vs eps (WV = c0 + c1 eps): A -0.85+-0.20, B -0.42+-0.21, corner +0.43+-0.20, unused -0.80+-0.23. The raw WV of the positive boxes rises as eps falls and the corner WV falls in magnitude with eps: an additive offset predicts exactly the first, the ideal circuit's finite-strength response the second. ## 5. 24 Sep - dark_floor_runs/raw_threebox_all.json holds per-shot s0, s1, p for the four jobs. Recomputed exactly: fez std corner(C) -0.498+-0.056; kingston -0.587+-0.057; marrakesh -0.689+-0.057; marrakesh corner variant -0.506+-0.057. The four are mutually inconsistent at about the 2 sigma level (chi2 about 7.3/3, p about 0.06); mean -0.57+-0.03. - Positives in the same runs: marrakesh std 0.579, 0.579; kingston 0.798, 0.843; fez 0.753, 0.909; marrakesh-11 0.773, 0.763. Null boxes (kept Xp) +0.026, +0.087, +0.043, +0.085 (mean +0.060+-0.016). - The saved FakeMarrakesh numbers are single 8192-shot draws (-0.829, -0.810, se 0.055); the exact model values, depending on transpiler layout, are -0.76 to -0.81 (marrakesh). - Which qubits the 24 Sep jobs used is not recorded (only the qasm of the first pub of daqpd5rt is on disk). ## Caveats of the referee - FakeFez is a calibration snapshot, not the calibration of 11 Oct. The conclusion "no excess over ordinary noise" is "no excess once an additive probe offset is allowed" and that offset is itself not in the model. The model-free statement is the one in the paragraph below. - Real hardware here is one chip, one day, 2 to 5 qubit triplet; single instances for eps 0.2, 0.15, 0.1. ## Paper paragraph (numbers only) On ibm_fez (qubits 142, 143, 144; 8192 shots per circuit; 9 jobs, 11 Oct 2026), the Aharonov-Vaidman three-box circuits with post-selection s0=s1=0 kept 12% to 18% of shots. The probe expectation divided by 2 eps was +0.82+-0.02 (A), +0.79+-0.02 (B), -0.48+-0.02 (corner), +0.04+-0.03 (unused) at eps=0.4, and +0.91+-0.04, +0.78+-0.04, -0.59+-0.04, +0.10+-0.04 at eps=0.25; at eps=0.1 the corner value was -0.60+-0.14. The noiseless transpiled circuits give +0.90, +0.90, -0.80, 0.00 at eps=0.4 and +0.96, +0.96, -0.91, 0.00 at 0.25. A fit = S + P sin(2 eps) over six strengths gave P = +0.82+-0.05 (A), +0.82+-0.06 (B), -0.53+-0.05 (corner), -0.10+-0.06 (unused), with S = +0.07, +0.03, -0.02, +0.11 (+-0.03 each); the same fit of the noiseless circuit gives P = +0.98, +0.98, -0.85, 0.00 and S = +0.01, +0.01, -0.04, 0.00. The unused box showed a mean of +0.062+-0.010 over all strengths (+0.041+-0.004 over all shots). A FakeFez noise simulation of the same circuits gave P = +0.80, +0.86, -0.67, +0.02 and S = +0.03, 0.00, -0.01, +0.01; adding one common offset of 0.05 and a gain of 0.96 to it reproduces the 24 hardware (box, eps) means with chi2 = 41.6 for 22 degrees of freedom. Duplicate jobs agreed (chi2 = 11.1 for 12). Single runs on 24 Sep at eps = 0.25 gave the negative box at -0.50+-0.06 (ibm_fez), -0.59+-0.06 (ibm_kingston), -0.69+-0.06 (ibm_marrakesh) and -0.51+-0.06 (ibm_marrakesh, corner variant).