In-loop RL calibration / control steering dec_rl

RL as a control optimiser on superconducting qubits is not new: the widely quoted 99.92% CZ record between two fluxonium qubits coupled through a transmon is an RL-optimised mean of 99.922(9)% from 2023 fact, but offline and per-gate. Sivak, Morvan, Broughton, Cortiñas et al. (Google, arXiv 2025-11-11; Nature 655, 2026-07-08) changed the loop: detection events serve as both syndromes and learning signal, so calibration runs concurrently with computation [2].
a = 1.0, a fully engineered classical-control layer, f = coherent — it chases systematic drift (TLS, flux, temperature), not stochastic Pauli error [graph].

No new physical floor: this tracks an existing floor without the dead time between calibrations. The load-bearing claim is architectural — optimisation speed independent of system size D[2], the only route past a per-parameter sweep that dies above ~10³ knobs. The limits follow from what the agent sees and what it turns: an averaged rate of detection events, and single-qubit and CZ gate parameters only D[2]. Willow's roughly hourly correlated bursts do produce detection events, but they decay within ~400 µs D[1], far faster than the agent can respond, and the paper leaves such fast correlated drift to hardware mitigation D[2]. Leakage out of the computational space, which Willow clears with a dedicated removal step after every round D[1], is not treated in the paper, and readout, at ~10⁻², is not among the tuned parameters. It does not change the exponent: no Λ is reported for the record d=7 run, so what moved is the point error at fixed distance, not the slope that decides whether d≈25 reaches 10⁻⁶.

The premise of in-loop recalibration is contested by a 2026 measurement: long-lived two-level systems (TLS) on a transmon give the environment 20 µs to 1 ms of memory — hundreds of ~1 µs error-correction cycles — so errors become history- and protocol-dependent, which the authors contrast with the Markovian (memoryless) assumption behind continuous calibration P[890].

AlphaQubit 2's own real-time benchmarks (arXiv:2512.07737) ran on recorded data, and the reinforcement-learning paper does not state that its d=7 decode ran live.

No dedicated silicon and, unlike matching or clustering decoders, no FPGA resource claim to audit fact. The cost is control-plane traffic — detection events out, parameter updates in, over the decoder's own link; NVQLink's 3.84 µs mean, 3.96 µs maximum round trip is the vendor-neutral version of it Cfact. Updates run on a slow timescale outside the 1.1 µs cycle, which is why it coexists with Willow where heavier in-loop computation would not. Supply is GPU/TPU capacity plus the control stack; no specific export rule.

Requires no decoder inside its control loop — the reward is the detection-event rate; its decoder steering reweights a matching graph from logical-error estimates, not in real time, and the record d=7 run was decoded separately with the neural decoder AlphaQubit 2, published on its own only as a December 2025 preprint whose real-time speed was benchmarked on recorded or simulated data D[2]fact. Superconducting transmons only; derived round unchanged at 0.65 µs against the measured 1.1 µs QEC cycle, ~0.91 MHz. Verification: no Λ for the record run, no second-platform replication, and the only live closed-loop neural decoding anywhere is a d=3 demonstration with a 550 ns loop D[729]. No value conflict: the 20% and >1,000 parameters in trade press P[742] are the paper's own — about 20% further logical-error suppression after conventional calibration, over a thousand steered parameters — beside its 3.5× stability gain against injected drift D[2].

Money.
2024-10-08 · Q-CTRL · Series B expanded · USD $113 M (AUD $166 M) · GP Bullhound lead · closed C[743]
2025-10-03 · Google Quantum AI · absorbed Atlantic Quantum, cold-stage control · undisclosed Pfact
2025-11-06 · DARPA QBI Stage B · IBM the only transmon vendor, Google absent · ≤$15 M each fact

Market & supply chain. No component market; the priced substitute is calibration software plus GPU/TPU cycles, sold only by Q-CTRL, on Diraq, Oxford Quantum Circuits and Rigetti C[743]. Pays G3/G4.

IP & standards. No dated patent family for RL-steered QEC calibration as of 4 Sep 2026; Google publishes the method and withholds the decoder — that is the moat.

Roadmaps & track record. Google's next milestone, a long-lived logical qubit at 10⁻⁶ (no date): 7.72×10⁻⁴ at d=7 is three orders short, needing d≈25 at Λ≈2 D[1]; Google is absent from QBI Stage B G. Q-CTRL shipped Boulder Opal on three platforms by 2024-10 — delivered, at nothing like Google's scale.

Strategic reading. Software margin on hardware someone else owns. If it diffuses, control vendors and Q-CTRL capture it and calibration headcount falls; if AlphaQubit 2 stays in-house, Google's advantage is the decoder, not the RL. Substitution threat: hardware that drifts less — fluxonium, cold-stage SFQ control — shrinks the problem.

Confirm if RL-in-loop calibration appears on a second platform, or a Λ is published for an RL-steered run; demote if it stays Willow-only. Best case 2029: standard across superconducting FT stacks, sold by control vendors. Worst case: one unreplicated chip. Open: does size-independent optimisation survive outside simulation, and can the loop be made decoder-agnostic?