Three things this time — a claim, a measurement, and two new tools.

IBM says “quantum advantage”. The NZZ says “not so fast”. Both are useful.

On July 30 IBM and partners (University of Chicago, Qedma, Algorithmiq) announced three simultaneous quantum-advantage demonstrations — verifiable sampling on about 70 error-corrected logical qubits, and two materials simulations that classical supercomputers could not match — and posted them to a public tracker with an open challenge to refute them (arXiv:2607.25941; criteria in arXiv:2506.20658). Two days later the NZZ ran the counter-view: Joseph Tindall of the Flatiron Institute, who advised one of the studies, called the material “90% artificially constructed” — a classically easy problem tuned until classical methods failed — and ETH’s Johannes Knörzer summed it up as “solid progress, not yet the big breakthrough” (NZZ, Aug 1). My reading: advantage is becoming real in physics; in finance it remains open, and none of the three demos touches it. The question for a bank was never “is there advantage” but “is there advantage on my problem, measured honestly” — vendor demos run on problems built for the machine. And IBM’s open tracker is the stance we take with the certified referee: publish, invite refutation.

So we measured our own tools the same way.

With a certified referee in the loop, every reasonable quantum method eventually finds the optimum on prototype-size portfolio problems — a 0% gap alone doesn’t rank methods. What does is concentration: how much probability the trained circuit puts on the proven optimum, and whether it is the single most likely readout. We ran 12,397 exact simulations across four method choices — training objective, circuit family, start state, penalty weights — on 20 synthetic portfolios and four real demo books, every run scored against exhaustive enumeration, design written down before the first run. Six findings, two of them null results, and the answer is a decision tree, not a method:

The 16-page study is public, pre-registered design and deviations included: QAOA Concentration Study (PDF) — condensed next to the sample runs at qubit-lab.ch/evidence.

Next month’s letter: what happened when we put these circuits on a real IBM quantum processor, measured against the same referee — the study is already on the evidence page, if you can’t wait.

New in the Rapid Quantum Prototypes (RQPs)

New RQP: Credit Allocation Pro

Batch credit decisions when exposures move as one — portfolio acquisitions, risk-transfer pools — under a capital budget, sector caps, and the fact that defaults cluster; the quadratic heart is correlated unexpected loss. On the shipped book, default settings failed honestly at 40.8% above the certified optimum; the tuned configuration reached it exactly, and the report published both. On the 23-qubit acquisition demo (8.4 million possible baskets) the referee proves the optimum in minutes — and the classical heuristic missed it by 0.19%.

New RQP: Robust Stochastic Collateral Allocation

Tuesday-morning margin calls: assign free collateral to venues at minimum posting cost while a scenario term keeps coverage stable under stress — locked lines, pledge pools that move whole, per-venue haircuts and FX, limits priced or hard. On the treasury sample the quantum run matched the certified optimum at 0.00% gap with the optimum as the single most likely readout of the register; the frontier prices coverage stability: roughly CHF 19,000 a year buys the coverage volatility down from 8.3M to 7.6M, certified point by point.

Genuinely new: a quantum circuit that builds itself

We ported the adaptive-ansatz idea from quantum chemistry — ADAPT-VQE, where the circuit grows one operator at a time from a pool, taking whichever helps most, and stops when nothing does — into our QAOA tools, trained on the CVaR tail of the readout. Its point is not speed (it costs several times a standard run) but independence: it does not inherit the classical seed, so it is the second opinion for the case where you doubt the incumbent answer. Together with the Egger-style classical seed start (in Credit and Collateral Allocation with a blur you set) it is now a switch in all three QAOA-based RQPs — Portfolio Optimization, Credit Allocation, Collateral Allocation — and every report prints the concentration next to the certified gap. The study above is what changed our defaults; the manuals say which and why.

The suite counts seven tools now — Portfolio Optimization, Index Replication, Feature Selection, Quantum Classification, Quantum Monte Carlo, and the two above. Two-page overview with real numbers: reply “overview”.

On LinkedIn, from this week

Two short series in parallel through early October: the story on Tuesdays — how a prototype works, this week FINMA’s PQC roadmap, then what an honest benchmark looks like and where quantum earns its keep — and one measured case every Friday (index replication, credit, collateral, QML — including the run where our own tool said “stay classical”). Rather have it here than there? Reply “digest” and I’ll bundle it into the September letter.

Quantum risk, briefly: if you missed last week’s note on FINMA’s July guidance (05/2026) — a board-adopted PQC roadmap recommended by mid-2027, and only 8% have one — it’s here: qubit-lab.ch/pqc-navigator. Different track, same principle: measure before you commit.

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