We Retracted Three of Our Own Quantum Claims. Here Is What Each One Got Wrong.

Updated 19 September 2026.

In May 2026 we retracted three quantum-hardware results we had published. An outside article on a related subject made us re-open our own raw data, and when we did, three results we had liked did not hold up. This post is the record of what each one claimed, why it was wrong, and what we assert now in its place. The corrected rows are on our claims page, with the job identifiers IBM issued wherever a single job isolates the claim.

An earlier, shorter version of this note appeared on this site. It said we had pre-registered these tests before running them. We had not, and that sentence is the first correction here.

A second correction came in September 2026, and it goes deeper. For several of the March runs below, the IBM job ran, but a bug in our code meant its results were never read back, and the numbers we published were computed on a local simulator instead. Each section below says where that applies.

First, the discipline we did not have

When these circuits ran, in March 2026, nothing had been written down in advance about what a positive result would look like or what null it would have to beat. We had data, we saw the shapes we hoped for, and we labelled them. That is the failure all three retractions share, and it is more instructive than any of the individual mistakes.

Pre-registration is how we work now. The success criterion and the null model are written before the hardware runs. We adopted that rule because of what follows, not before it, and the record should say so.

1. The fine-structure constant "detection"

What we published. A 15-qubit phase-estimation circuit, 128 measurement bins, 8,192 shots, submitted to IBM hardware. The write-up claimed a signal-to-noise ratio of 25.4 at a phase angle of pi over 137, and read that as the fine-structure constant showing up in hardware.

What was wrong. Two things, and the second was worse. The ratio divided each bin's probability by a flat floor of 1 in 128, as though every bin were equally likely under the null hypothesis. The eight brightest bins turned out to be the eight nearest phase zero, in a histogram mirror-symmetric around it: a pile-up at zero phase, not an eigenphase at a chosen angle, with the label attached after the fact. Then, in September, we found that the job had run on an older 133-qubit IBM machine but its results were never read back. The histogram we analysed was computed on a local simulator. There was no hardware data in the claim at all.

A second artifact made the same class of error in a cruder way. An 8-qubit structured quantum walk had recorded "alpha detected: true" while its list of detection hits was empty. A detection flag with no detections behind it is a label, not a result. Its numbers turned out to be simulator output as well.

What we say now. The jobs ran on IBM hardware. No hardware result from them exists in our analysis, and no eigenphase was resolved in the numbers we did have. The angle pi over 137 remains something we chose as an input when designing circuits, and we describe it as that: a design parameter, not a measured constant. The IBM job identifier for the phase-estimation run is d6ts6i0v5rlc73f38ftg.

2. Shor's algorithm and the discrete logarithm

What we published. Shor-style circuits for the elliptic-curve discrete logarithm executed on IBM hardware and recovered private keys.

Precision first. These were ten small prime-order curves, with group orders from 4 bits to 21 bits. No production curve was ever targeted. Any wording of ours that implied otherwise, including in the earlier version of this note, was wrong.

What was wrong. Keys did come out of the pipeline. They came from its classical half. At these sizes, direct classical enumeration finds the key by itself, and the quantum measurement data played no causal role in the answer. In the hardware results, read on their own, the correct key sits at the noise floor.

What we say now. Twelve jobs ran on IBM hardware, using up to 41 qubits, eleven of them on ibm_fez. For eight of the ten curves, the results we analysed came from that hardware. For the 6-bit and 8-bit curves they did not: the same code bug meant the published results were computed on a simulator. Zero of the ten support a quantum break, whichever source the numbers came from. The job ledger is the whole claim, and nothing more than that. Two of the identifiers, for anyone comparing against the ledger: d6ue2rk69uic73ci52f0 and d6ue2kif84ks73de5o1g.

3. The "Bell-certified" 16-qubit circuit

What we published. A 16-qubit entanglement circuit reported with "Bell fidelity 1.0000", and downstream of it a keyed HMAC described as seeded by a "Bell-certified gate".

What was wrong. A fidelity of exactly one on real hardware should have been the alarm by itself. When we went back in September to find the run, we found no record of it in our run log and no IBM job we can attribute to it. The certification language had no measurement behind it, and the HMAC framing inherited the problem.

What we say now. We no longer claim this circuit ran on hardware, and we claim no result from it. No Bell certification is claimed. The HMAC stands as a keyed HMAC; we no longer describe its entropy as coming from hardware.

What a real pass looks like. A later CHSH test, run properly on ibm_marrakesh, gave S = 2.733 with an uncertainty of 0.023, which is 32 standard deviations above the classical bound of 2. It is a reproduction of textbook physics on a single device, with the locality and detection loopholes open, and our claims page lists it as exactly that. The distance between an unsupported "fidelity 1.0000" and a device's "2.733, plus or minus 0.023, loopholes open" is the distance between a label and a measurement.

Why publish the failures

Anyone can publish the runs that flatter them. What is left after we removed ours is a claims page where every row carries a status, a floor, and a receipt. The vocabulary is small on purpose: executed, corrected, unresolved, reproduction. Where a claim overreached, the row says so and the correction is the receipt.

One honest limit on those receipts. IBM job identifiers are issued by IBM and belong to the account that ran the jobs. They are a record we cannot edit, which is why we list them. They are not a button a stranger can press to replay the run. We anchor our claims to build numbers and dates, and we publish a fixed procedure for overturning any of them.

The three retractions went out on 15 May 2026. Everything we still assert about quantum hardware is on the claims page, at the floor the data supports and no higher.