In quantum, the verifier is now the story

Facts are attributed and linked to their sources. The analysis and opinions are the author's own.

What it is

On 7 October IBM announced that it had been selected for Stage C of DARPA's Quantum Benchmarking Initiative (QBI). IBM describes Stage C as the phase where DARPA's independent verification and validation team moves from reviewing plans to testing companies' hardware, to determine whether a company can build a fault-tolerant quantum computer. Earlier stages asked for a technical concept (Stage A) and a detailed R&D plan with risks (Stage B). IBM Research director Jay Gambetta said IBM "expects to deliver the first fault-tolerant system in 2029".

Five days later, Quantum Computing Report described a Google Quantum AI preprint on Virtual Error Cancellation, a classical post-processing method that uses recorded syndrome data and two decoders, a fast real-time one and an accurate offline one, to reweight experimental shots and cancel residual logical errors. In simulations on a distance-7 surface code, the report says, residual logical errors fell by more than 1,600-fold versus unmitigated decoding. These are simulation results, not a hardware demonstration.

Why it matters

Quantum computing has suffered from a measurement problem for years: every vendor picks its own metric, qubit counts are quoted without error rates, and 'advantage' is claimed on problems chosen to show it. An external programme whose explicit job is to test hardware against a fault-tolerance goal is the most important development here, more than any single qubit number.

A metric reported by the team being measured is a hypothesis. A metric checked by someone independent is evidence.

This is the same lesson I keep returning to in analytics. A metric reported by the team being measured is a hypothesis; a metric checked by someone independent is evidence.

The structure of the programme is worth borrowing. Stage A asks for a concept, Stage B for a plan with risks, Stage C for test results. That is roughly how a good model-risk process treats a new credit model: design document, development evidence with known limitations, then independent validation before use. The order matters, because each stage filters out ideas that would be expensive to discover as failures later.

For readers outside physics, the key distinction is between physical and logical qubits. Physical qubits are noisy. Error correction combines many of them into one logical qubit that is more reliable, but only if the physical error rate is below a threshold; above it, adding more qubits makes things worse. That is why error rates, not counts, are the number to watch. The interactive explainer in this section lets you see that threshold behaviour for yourself, with stylised numbers.

The Google work points at another familiar idea: use a slow, accurate check after the fact to correct a fast, approximate decision made in real time. In fraud and credit operations we do this constantly, scoring quickly at the point of transaction and reviewing more carefully afterwards. The quantum version is mathematically very different, but the operational logic is the same.

What I'd watch

• What the independent testers publish. A pass or fail is useful; the measured logical error rates and the test conditions would be far more so.
• Whether 2029 targets survive contact with Stage C testing. Dates are plans until verified hardware backs them.
• Software-side error handling moving from simulation to hardware, and how its overhead is reported.

Sources

1. IBM Advances to Stage C of DARPA Quantum Benchmarking Initiative. IBM Newsroom, 7 Oct 2026.
2. Google Quantum AI Introduces Virtual Error Cancellation for Universal Logical Circuits. Quantum Computing Report, 12 Oct 2026.