What is lattice surgery, and what was demonstrated on IBM hardware?

Lattice surgery is the standard way two error-corrected logical qubits interact on a two-dimensional chip: neighboring code patches are merged by measuring a joint logical operator, then split apart — turning protected memory into protected computation. The operation was proposed on paper in 2012 (Horsman, Fowler, Devitt, Van Meter, arXiv:1111.4022). As of November 6, 2026, the strongest demonstration on record is a logical Bell state between two distance-3 surface-code patches on IBM's cloud-deployed, 120-qubit ibm_phoenix processor, with fidelity bounded above 93% — a post-selected bound that kept 40% ± 2% of shots, reported by University of Sydney researchers (arXiv:2610.32347, 23 Oct 2026). Rosetta Quantum reads that number the way it reads every record: together with its accounting.

What is lattice surgery?

Lattice surgery is a method for performing logical operations between error-corrected qubits by temporarily merging two code patches along a shared boundary and then splitting them apart again. It was proposed by Horsman, Fowler, Devitt and Van Meter in a paper posted in November 2011 and published in 2012 (New Journal of Physics 14, 123011; arXiv:1111.4022).

The background, in one paragraph: a logical qubit in a surface code is not a device — it is a patch of physical qubits whose collective state encodes one protected qubit (we compared how vendors demonstrate them in our logical-qubit demonstrations post). Storing a logical qubit is a memory experiment. Computing requires logical qubits to interact — and on a flat chip where each qubit talks only to its nearest neighbors, you cannot lift one patch and carry it to another.

Lattice surgery solves this with three moves. A merge stitches the boundary between two adjacent patches by switching on stabilizer measurements across the seam for a number of error-correction rounds; the result measures a joint logical operator (such as X̄X̄ or Z̄Z̄) of the two encoded qubits without measuring either qubit individually. A split switches those seam measurements back off, restoring two separate patches. Composing merges and splits — with a small ancilla patch where needed — implements a logical CNOT, state teleportation, and with magic states the rest of a universal gate set. Every step uses only nearest-neighbor operations on a plane, which is precisely what superconducting hardware offers. The same «measure a joint operator, then separate» idea generalizes beyond surface codes: in October 2026 a third-party team demonstrated code surgery on a [[10,2,3]] quantum LDPC code on Quantinuum's H2-2 (arXiv:2610.39067), which we covered in the Quantinuum record explainer.

Why do logical qubits need surgery at all?

Because a logical qubit that only preserves a state is a memory, and a computer is not a warehouse. Every fault-tolerant algorithm needs logical two-qubit operations, and lattice surgery is the lowest-overhead known way to get them on 2D hardware with nearest-neighbor connectivity: the influential compilation scheme for large-scale surface-code machines expresses entire algorithms as choreographed sequences of patch merges and splits (Litinski, «A Game of Surface Codes», Quantum 3, 128, 2019).

Two consequences follow. First, the merge is where the classical decoder earns its pay: joint measurements must be decoded across the merged patch, round by round, and feed-forward decisions eventually have to arrive faster than the hardware cycles — sub-microsecond, in the Sydney authors' own framing of what remains (arXiv:2610.32347). Second, surgery alone gives Clifford operations; the non-Clifford gates that make a quantum computer more than classically simulable still require magic states, whose cost we broke down in our magic-state distillation explainer.

What was demonstrated on IBM hardware in October 2026?

The headline result: a logical Bell state — the minimal unit of logical entanglement — prepared via lattice surgery between two distance-3 surface-code patches, with fidelity bounded above 93%, plus logical state and gate teleportation across three logical qubits at 78% to 93% process fidelity (arXiv:2610.32347, posted 23 Oct 2026).

The experiment was run by University of Sydney researchers (Harper, Kolesnikow, McLauchlan, Nixon, Bartlett, Brown) on ibm_phoenix, a 120-qubit square-lattice Nighthawk r2 processor that entered IBM's cloud fleet in early September 2026 (Quantum Computing Report, 2 Sep 2026). The preprint landed 54 days after the processor's fleet integration (Quantum Computing Report, 28 Oct 2026) — an academic team, on commercial cloud hardware, at record fidelity. That provenance is part of the story: this was not a hero apparatus in a single lab.

Two hardware ingredients did heavy lifting. Nighthawk r2's dissipative reset couples each qubit to a cold bath on demand, cutting its effective relaxation time from ≈200 µs to ≈25 ns (Quantum Computing Report, 2 Sep 2026) — fast, clean resets between rounds. On top of that, the team ran a «walking» syndrome circuit that alternates the data and ancilla roles of the physical qubits each round, so every qubit is measured and reset every two rounds — a mechanism aimed squarely at leakage, the slow poisoning of a code by qubits stuck outside the computational states (arXiv:2610.32347; QCR, 28 Oct 2026).

What does «above 93%» actually certify?

It certifies a lower bound on logical Bell-state fidelity, under an accounting that the paper declares and a reader should repeat. Four entries matter (all from arXiv:2610.32347):

The number is a bound, not a point estimate: it is computed from the measured ⟨X̄X̄⟩ and ⟨Z̄Z̄⟩ correlators because ⟨ȲȲ⟩ could not be reliably measured. The shots are filtered: only runs confidently initialized inside the code space are kept, a post-selection with a yield of 40% ± 2% — roughly 60% of the 500,000 shots per circuit are discarded before the fidelity is computed. The chip is curated: one defective qubit (Q55) is excluded by re-purposing an ancilla, and the paper states that defect exclusion was needed to reach the >93% level. The decoding is model-informed: the workhorse decoder is belief matching with a detailed noise model, stated explicitly for the memory experiments.

None of this is cheating — it is all declared, which is exactly what honest benchmarking looks like. But the headline number describes the post-selected, defect-excluded performance of the logical layer, not the unconditional performance of the machine. A 40% yield is a real cost: a fault-tolerant algorithm cannot throw away 60% of its runs at every operation. This is the same lesson Rosetta keeps finding in GHZ records and logical-qubit demos: a number travels with its accounting, or it means less than it seems.

How do the three lattice-surgery Bell states compare?

Measured rung by rung, the ladder is short — three Bell states in fourteen years of the idea existing (fifteen since the arXiv posting):

The first demonstration (Innsbruck, 2021) entangled two distance-2 patches of four ions each and reported 58.0(1.6)% raw — above the 50% separability limit by 5 sigma, which is what made it a result (Nature 589, 220 (2021)). The June 2026 Zhejiang experiment reached d=3 on a 125-qubit superconducting chip, but its raw surgery fidelity was 0.372 — below the entanglement threshold — rescued to ≥0.532 by decoding and to ≥0.925 only by keeping ≈0.5% of runs (arXiv:2606.06598).

That makes the sharpest comparison in this table not the fidelity but the yield. At comparable post-selected fidelity (≈0.92–0.93), the kept fraction went from ≈0.5% in June to 40% in October — roughly eighty times more of the data surviving the filter, with the fidelity bound intact. The caveat is the same one Rosetta attaches to every cross-vendor table: no two rows share an accounting regime (raw vs. decoded vs. post-selected, error-detection vs. error-correction codes, trapped ions vs. superconductors), so the ladder shows direction, not a ranking.

What is still missing between this demo and fault-tolerant computing?

The authors say it themselves: «we require a larger number of logical qubits performing with much better fidelity than those we have demonstrated here» (arXiv:2610.32347). Measured against that destination, the open items are concrete. Real-time decoding will have to make feed-forward decisions in sub-microsecond times; the published results are decoded after the fact. Surgery will have to run at distance 5 and beyond — the demonstrated merges are distance-3, and the authors flag residual crosstalk on dense square lattices as the thing that will have to be engineered down. Non-Clifford operations — the magic-state layer — will have to be added and benchmarked; this demo is Clifford-only (our magic-state explainer covers why that layer dominates cost). And the scale gap is two orders of magnitude: two or three logical qubits exist on this chip today; the authors' long-term target for industry-scale algorithms is logical error rates of 10⁻⁸ or below, with hundreds of logical qubits. The paper itself points toward larger arrays, longer-range couplers and qLDPC codes as the path (arXiv:2610.32347) — the same direction as the cross-code surgery already shown on H2-2 (arXiv:2610.39067).

None of that diminishes what was measured. It locates it: lattice surgery has moved from a 2012 proposal to a >93% logical Bell bound on a rentable cloud device in fourteen years — and the distance from here to fault-tolerant computation is still measured in missing layers, not missing percentage points.

What we don't know

A point estimate with error bars for the Bell fidelity: the abstract states «above 93%»; no central value with uncertainty is quoted in the text we reviewed. The raw, un-post-selected Bell fidelity: the paper shows it as a marker in a figure but does not quote it as a number in the text we reviewed. Which decoder produced the Bell bound: belief matching is stated for the memory experiments; its role in the Bell result is not stated with certainty. Peer-review status: none — the preprint was 14 days old on our publication date, and peer review will take months. Replication: none on another device of the same class as of November 6, 2026. Scaling of the walking circuit: whether the leakage-flushing approach holds at distance-5 surgery under crosstalk has not been measured; the authors list it as future work.

Rosetta Quantum publishes verdicts with raw, reproducible data. This is educational content — not a product claim, and not investment advice.