Closed-loop AI needs a verifier coverage map
In short: Anthropic's quantum-circuit search shows that closed-loop AI is defensible only across the dimensions its evaluator can verify.
On 16 October 2026, Anthropic described a Claude Mythos 5 search across eight quantum-circuit design tasks. An independent program checked every proposal for correctness and scored its qubit and gate cost; the accompanying paper reports improvements on seven tasks.
That result does not prove that long-running AI loops are broadly safe. It shows something more useful: closed-loop AI needs a verifier coverage map. Autonomy is defensible only across the dimensions the evaluator can actually check.
Anthropic's quantum-circuit loop had a narrow judge
The paper submitted on 15 October describes an evolutionary loop in which an LLM proposes circuit-generating programs and an independent evaluator verifies and scores them. Quantum circuits were a strong fit because their outputs could be checked on ordinary computers, while qubit count multiplied by gate count supplied an objective optimisation target.
Anthropic reports that the discovered circuits reduced both Toffoli-gate and qubit counts against prior designs in seven of eight computational primitives. For the eighth, the search produced a new reversible architecture for modular inversion.
| What the loop checked | Reported evidence |
|---|---|
| Search duration for the elliptic-curve design | 53 hours |
| Simulated Reed-Solomon decodings | 403 million |
| Overflow failures in that decoder design | about 1 in 10,000 |
| Gate reduction versus Google's earlier decoder | 12x |
These are author-reported research results from Anthropic and the preprint; I did not independently reproduce them.
The evaluator did real work. It rejected incorrect candidates, exposed trade-offs and returned a score the model could improve. But it did not prove production quantum advantage, hardware feasibility, research importance or that the chosen cost function captured every valuable design property.
A verifier coverage map sets the loop's boundary
My rule is to describe every acceptance dimension as one of three things: deterministically checked, statistically sampled or human-reviewed. Anything outside those columns is not covered merely because the loop keeps producing higher scores.
That is the verifier coverage map. For a coding agent, compilation may be deterministic, regression tests may be sampled, and architecture fit may remain human-reviewed. For a support agent, schema validity may be deterministic, resolution quality sampled, and an irreversible customer concession human-owned.
The map prevents a common category error. A verifier can establish that the system passed its gate. It cannot silently inherit responsibility for the dimensions nobody encoded.
My practitioner correction is to budget the uncovered surface
This is editorial judgment, not a claim that I ran Anthropic's quantum experiments at Trigent or for a client. In enterprise agent work, I would permit another autonomous iteration only when the current state is recoverable and every consequential change lands inside an explicit acceptance path.
The operating record should name the candidate artifact, evaluator version, dimensions checked, test distribution, known false-accept and false-reject modes, uncovered dimensions, rollback point and human owner. If the agent changes the evaluator, environment or objective, the old certificate expires.
This advances an earlier archive distinction. Astra's formal verification did not prove it discovered the proof, while an adaptive agent environment can change the path without changing the final verifier. Here the question comes before both: what portion of the task is verifiable enough to enter a closed loop at all?
Where closed-loop AI still needs human review
Anthropic's own account says the model sometimes preferred familiar approaches and needed occasional prompts to broaden its search. The 53-hour run therefore had minimal human input, not zero human influence. Its strongest result came from a task where correctness and resource cost were unusually legible.
Most enterprise work is messier. A green test suite can miss a contract change. A high resolution score can hide an unfair exception. A cheaper workflow can move risk onto a reviewer. Closed-loop iteration remains useful, but the uncovered surface needs a budget and an owner.
What's in it for you
- Mark every acceptance dimension as deterministic, sampled, human-reviewed or uncovered.
- Stop the loop when its evaluator, environment or objective changes without re-qualification.
- Report improvement only inside the verifier's scope, then name the remaining human decision.
Do not ask how long an AI can run alone; ask how much of its work the verifier can truly see.
