Three stories this month: real progress on reliable qubits, our circuits on a real IBM processor, and code exports across the whole suite.

The month’s quantum news: real progress on the road to reliable qubits.

Some background: quantum bits are fragile, so a dependable “logical qubit” is normally built from many physical ones. The textbook expectation is hundreds to thousands of physical qubits for one reliable logical qubit, and that overhead is the main reason useful machines are said to be years away. This month brought real movement on that number. On September 24 the US company Infleqtion entangled 30 logical qubits built from just 80 physical atoms on its commercial neutral-atom machine, less than three physical qubits per logical one (technical write-up). The fine print (independent analysis): these qubits notice errors and discard the affected run rather than fixing them mid-computation. A smoke detector rather than a sprinkler, in other words; the error-fixing kind remains the next milestone on the roadmap. Still a real step: the previous record for entangled logical qubits was 24, on a much larger machine. Thirty from 80 atoms is new, and scaling up cheap detection is a recognised stepping stone toward full correction, which so far exists only on single logical qubits. The same day Pasqal reported trapping 1,024 defect-free atoms and a first full application run on protected qubits (H1 results). My reading: the hardware curve is bending, and if it keeps this pace, timelines shorten.

As promised: our circuits on a real IBM quantum processor.

Last month’s study ranked quantum methods by concentration (how much probability the trained circuit puts on the certified optimum) in exact simulation. This month we took five of those trained circuits (depths 1–7 plus the adaptive one) to ibm_kingston, one of IBM’s Heron processors: 20,000 measurements per circuit, 44 seconds of processor time in total, scored against the same referee. Three findings:

The practical reading: today’s hardware rewards shallow circuits and honest measurement, and concentration remains a simulator asset until fidelities improve. That is why our reports print both. The full study, design and deviations included, is public: Noise Ladder Study (PDF), condensed at qubit-lab.ch/evidence.

New in the RQPs: export the code. Test before you port.

“If we test with your tool, are we locked into your tool?” As of this month, the opposite. Every Rapid Quantum Prototype now exports what actually ran, the exact circuit, the trained parameters and the terms of your workbook, as runnable notebooks for Qiskit, PennyLane, Google Cirq, NVIDIA CUDA-Q and Classiq Qmod, plus the raw circuit as OpenQASM 2.0. Two flavors, because two audiences: rerun reproduces the certified result exactly (the auditors’ version), retrain opens the full training loop for your quants to continue in their own stack (the researchers’ version). The evidence travels with the code: the exported notebook carries the same recipe that produced the certified report, so porting doesn’t orphan the benchmark.

Verified where we can prove it: the exported models of all seven RQPs were synthesized and executed in Classiq’s own Studio environment, including a 23-qubit sample run; CUDA-Q exports are verified at the gate level against the shipped circuits, with runtime execution in your environment. Standard .ipynb and .py files, no license server, no phone-home. Sample notebooks sit on every product page and in the resource room. And every report across the suite now opens verdict-first: the answer on page one, the evidence after. Framework names are their owners’ trademarks; no partnership implied.

Two configuration atlases, published.

September’s science: two systematic maps of quantum machine learning configurations, both public with DOIs. The Configuration Atlas of Quantum Kernels measured 150 quantum-kernel configurations and answers which settings can be measured at all on real data. There is a predictable workable corner at every size, and it is smaller than the literature suggests. The Configuration Atlas of Variational Classifiers answers which settings can be trained: most of them, if you measure locally, and a one-minute screen tells you in advance. Both maps now run inside the QML prototype as guardrails, and the Evidence Engine behind these letters counts six published studies.

In Basel, and on LinkedIn.

September also took me to the QAI Ventures Quantum Global Summit in Basel: a day of good conversations with banks, quantum vendors and researchers, and the reason several of you are reading this letter for the first time. Welcome. On LinkedIn, Tuesdays now carry the study series (how these studies are made, then one study a week in management language) and Fridays the measured cases, most recently the self-building circuit that lost, won, broke and shipped, in that order. It runs through November; if LinkedIn isn’t your place, everything lands on the evidence page anyway.

Corporate networks: if links are blocked at the desk, everything here also works from a private device.

If any of this is worth a conversation, just reply. Happy to show everything live, your own workbook included. And if this lands closer to a colleague’s desk than yours, feel free to forward it.

Daniel

Daniel Hug · qubit-lab.ch · daniel.hug@qubit-lab.ch