Measured,
on record.

BELL  Distance-2 detection code, post-selected (~96% acceptance); the textbook p→p² suppression, so we quote no improvement factor. The measured correlators coincide with the stabilizers of one native coupler pair, and the circuit is Clifford — classically simulable. Not error correction, no QEC cycles, not unconditional, no logical gate, no real-time loop. Honest label: encoded + detected. Encoded [[4,2,2]] Bell states are published prior work on several platforms, including this one; we claim no first. Raw per-shot arrays are archived and the number re-derives offline from them; no third party has verified this result — the circuit is depth 7 and costs nothing to re-run. Job ID d9npkuuij12s73fu66qg. ↩

FLEET  A fleet of 31 independent pairs, not one 62-qubit entangled state. Distance-2 detection code, post-selected: mean acceptance 92.1%, joint acceptance across all 31 blocks 9.4% (Z) / 6.0% (X) — post-selection does not scale, which is precisely why detection is not correction at scale. 31/31 blocks above the 0.5 bound with full confidence intervals clearing it (raw-array re-decode, 2026-08-15; mean witness 0.9964, min 0.9790). Heron-class hardware. Job ID d9mf6308csec73fa3o7g. ↩

LADDER  One-basis (bit-flip): a repetition code protects a classical bit, not a qubit, and can never be a fault-tolerant qubit. One measured suppression step to a ~1×10⁻⁵ per-cycle floor: the d=5 and d=7 rates are statistically indistinguishable, so this is not a below-threshold ladder. Rounds = d, so state preparation and measurement are not separated from the per-round rate. Our software refuses to quote a suppression ratio with fewer than 5 events. Heron-class hardware. No quantum advantage is claimed. Job ID d9nppncsfqic73arb48g. ↩

TAX  A measured cost, published deliberately: a statement about lattice connectivity, not a ranking of anyone’s device. Published below-threshold demonstrations run on degree-4 lattices; we ran on degree-3. At distance 5 the decoded error rate is statistically indistinguishable from chance in both bases, so no suppression ratio is measurable on this lattice — an earlier per-ratio figure from these runs was retired on 2026-08-15 after a raw-array re-decode found a clamp artifact, and this page reports the re-decoded numbers. The quotable distance-3 pair: Z-basis logical error 0.152 per cycle, X-basis 0.305 (marginal). Routed two-qubit gates inflate 5.2–5.4× per basis at distance 5. SWAP routing inflates the two-qubit gate count several-fold with distance; our transpilation was not connectivity-optimized. The simulation figure is simulation — it never sits on the same axis as the hardware number. Newer superconducting roadmaps are moving to higher-connectivity lattices; this measurement quantifies what the missing connectivity costs. This is the question our chip design starts from. Job ID d9lol8rhdfks73cl9m2g. ↩

LAYOUT  Compiler and transpilation metrics on heavy-hex topology — gate-count and routing facts about circuit compilation, not device-performance or fidelity claims. Our own before/after against the default toolchain, not a competitor comparison. Two-qubit count 3,710 → 615 (−83%) on the flagship circuit against the default toolchain, from the archived manifests. Placement quality self-validates: the block our placement ranked last was also the lowest-witness block. ↩

CONTROL  One-basis leg only; post-selected detection at distance 2 — not correction. The v1 X-basis run failed on a translation bug and is preserved, marked do-not-cite, rather than deleted. Detection gains: raw 0.563 → 0.705, reproduced 0.497 → 0.579. Points, not factors — the headline +0.14 is the raw detection gain. ↩

CLASSICAL  Classical, over GF(2), rate 0.229; minimum distance verified by exhaustive enumeration over all 65,535 messages. Honesty note: the Griesmer bound allows [35,16,12], so this construction is about twice as long as optimal — correct and tested, not record-setting. ↩

DECODER  Google Quantum AI’s measurement on Google’s hardware, not ours, and not pre-registered: the public shot record of its below-threshold surface-code experiment (Zenodo, doi:10.5281/zenodo.13273331, CC-BY-4.0). No job ID or acceptance rate applies: no hardware of ours, and every shot is counted. We re-decoded 17 of the archive’s 420 configurations offline (one patch each at distance 3, 5 and 7; 50,000 shots each) with PyMatching, open-source software we did not write, on the noise model in the circuit Google ships. Plain matching gives eight suppression factors from 1.23 to 1.90, all above 1: eight comparisons (distance 3→5 and 5→7, in the X and Z bases, at 13 and 30 rounds), not eight configurations, and not a reproduction of the published 2.14 ± 0.02. On the six 13-round configurations it makes 1.36–2.45× the logical errors of Google’s best shipped decoder (1.22–2.28× at 30 rounds). Others have re-decoded this record more completely; we claim no first. The home page’s model figure is a separate simulation of Google’s processor, built from its published T1, T2 and cycle time: not a simulation of our design, not a decode and not blind. One parameter is fitted to Google’s published distance-7 error rate, each distance is run for as many rounds as its distance (3, 5 and 7), and the archived unseeded run gives 2.11 ± 0.03 (sampling error of that one run) beside the published 2.14. Corrected 2026-10-13: this site earlier said eight “configurations” were “reproduced by our decoder”, and showed the model figure as 2.12, a “blind prediction”. Not real-time, not peer reviewed, no third-party check. ↩