Daily Dossier

OPEN — Retune while the music keeps playing

Quantum chips drift out of tune. Usually you stop the show to retune. Google taught an AI to retune from the same error beeps the computer already listens to — so the music can keep playing.

What changed

RL steers >1,000 Willow control params from QEC syndrome rates during computation — calibration no longer stop-the-world. ~20% LER cut vs expert calib; up to 3.5× stability under injected drift (with decoder steering).

Numbers

• Surface d=7 LER 7.72×10⁻⁴/cycle
• Color d=5 8.19×10⁻³
• ~20% post-expert RL fine-tune
• 2.4× control-only / 3.5× with decoder under injected drift

HOLDs / does not mean

• HOLD Not a new code
• HOLD Injected drift ≠ all natural drift
• HOLD “Never stops computing” is paradigm, not a months-long algorithm run

Hype reject

• “AI solves quantum error correction”

Primary: Nature 655, 879–884 (8 Jul 2026)

OPEN — Two protected qubits in 20, cheap Clifford logic

Protecting quantum info usually costs many spare qubits. Quantinuum’s Helix packs two protected qubits into 20 physical ones and can do the “easy” (Clifford) logic cheaply — and talk to a different code family that might supply the hard ingredients.

What changed

Early-FT architecture on Helios (not memory-only): [[20,2,6]] C₄-Helix memory + 2-logical Clifford RB under active correction + heterogeneous GHZ to d=5 surface code — beating physical baselines without postselection.

Numbers

• Memory ε₆ ≈ 4.6×10⁻⁵/cycle
• Logical Clifford ε_C ≈ 2.8×10⁻⁴ vs physical ~1.2×10⁻³
• GHZ LB ~99.925% vs physical ~99.54%

HOLDs / does not mean

• HOLD No T/magic-state injection
• HOLD Not universal FT
• HOLD Company preprint; thin independent media
• HOLD 10⁻⁶–10⁻⁸ targets are sims

Hype reject

• Naked “breakthrough”
• “Universal FT computer demonstrated”
• “3.5× better than surface” without metric

Primary: arXiv:2609.03194 — submitted 2 Sep 2026

OPEN — Moving and entangling protected patches, not just parking them

Instead of only parking information in protected memory, USTC/Pan moved and entangled error-corrected “patches” on a 107-qubit chip using surgery-like operations — the toolkit for computation, not just storage — though the gate quality is still modest.

What changed

No-postselection Clifford layer (H, S, CNOT) via lattice-surgery/deformation on d=3 surface-code patches with multi-round syndrome extraction + neural decoding — memory → active compute primitives.

Numbers

• 107 physical qubits · d=3 · no post-selection
• Logical CNOT F 0.643(2)
• Bell F 0.550–0.624

HOLDs / does not mean

• HOLD Not high-fidelity FT gates
• HOLD Not below-threshold
• HOLD Not universal (no non-Clifford without magic)

Hype reject

• QuantumZeitgeist “92.7% logical CNOT” (wrong)
• “Fault-tolerant processor”
• Conflating with postselected magic-gate papers

Primary: arXiv:2607.01473 — 1 Jul 2026

OPEN — Error correction on a chip that isn’t a textbook grid

IBM chips aren’t wired like a textbook surface-code grid. This USC/Lidar team still ran error correction on them by clever routing and noise-canceling pulses, showed that making the code taller or wider helps protect specific kinds of information, and warned that some popular scorekeeping methods can fake success.

What changed

Anisotropic surface-code scaling on IBM Heron heavy-hex (not isotropic below-threshold Λ): fold–unfold SWAP embedding + gap-aware DD; directional protection (3,5)/(5,3); DD kills spurious subthreshold claims; EF metric exposes misleading Λ fits.

Numbers

• 37-qubit d=3 → 65-qubit (3,5) and (5,3) embeddings
• On 156-qubit Heron (6 short of isotropic 5,5)
• ~30% noise cut → (5,5) within reach (authors’ estimate)

HOLDs / does not mean

• HOLD Not Google-style isotropic below-threshold
• HOLD ~30% is projection
• HOLD Not logical gates/algorithms

Hype reject

• “IBM surface code below threshold”
• “Subthreshold scaling” without anisotropic/EF caveats

Primary: Nat Comm, 29 Jul 2026 (arXiv:2510.18847)

OPEN — Continuous twist on an encoded qubit

Error-corrected qubits usually only get a small menu of exact turns (like 90°). This IonQ Forte experiment shows you can dial a continuous twist on an encoded qubit by twisting all the physical pieces the same way, then checking for errors — a baby step toward the fancy gates real algorithms need.

What changed

Proof-of-principle continuously tunable non-Clifford logical Z on [[7,1,3]] Steane via transversal RZ(θ) + error correction (not just discrete Clifford logical gates).

Numbers

• Code [[7,1,3]] Steane on Forte
• Transversal θ + EC → syndrome-dependent logical Z
• Two-round ±θ cancellation on trivial syndromes (cite paper figures for fidelities — no invented %)

HOLDs / does not mean

• HOLD Not FT universality at scale
• HOLD Not magic-state factory
• HOLD Not distance scaling
• HOLD Not production continuous SU(2)

Hype reject

• “Universal FT gates”
• “Continuous logical computing”
• Any algorithm speedup claim

Primary: arXiv:2608.20676 — submitted 21 Aug 2026

OPEN — Visible mistakes beat invisible ones

Most quantum errors hide. These qubits usually “lose the photon,” which the machine can see. D-Wave + Yale built a fast two-qubit gate that still mostly makes visible mistakes instead of invisible ones — which makes fixing mistakes much easier later.

What changed

First hardware-validated ~500 ns CZ on superconducting dual-rail cavity erasure qubits that keeps the erasure-over-Pauli hierarchy during two-qubit ops (not just at rest).

Numbers

HOLDs / does not mean

• HOLD Not a logical qubit
• HOLD Not below-threshold QEC
• HOLD Not quantum advantage
• HOLD Sims ≠ experimental QEC cycles

Hype reject

• “99.9% fidelity” without separating erasure vs Pauli
• “FT computer demonstrated”
• “solves QEC overhead”

Primary: Nature, 5 Aug 2026 (DOI 10.1038/s41586-026-10822-y)