definition
what is certified computation for AI agents?
Certified computation means an AI agent's numerical answer arrives with a certificate: a typed, machine-checkable statement of what was proven, by which verified instrument, to what tolerance. Not a confidence score, not a self-assessment. A contract beside the number.
The phrase also travels as certified mathematical computation and verified computation for agents. All of them name the same upgrade: the difference between an agent that produces numbers and an agent whose numbers can be acted on without a human re-deriving them.
the gap it closes
why agents need it
Today an agent asked to analyze data writes its own code inside a tool call: a dozen lines over your data, written once, gated never, trusted anyway.
The number that code returns looks exactly like a correct number would. Nothing visibly separates an answer that was proven from an answer that was merely produced, and the gap is paid later by whoever trusted it. Harnesses solved the permission problem: what an agent may touch. Certified computation solves the next one: whether what it computed deserves trust.
route, don't improvise
The agent hands over a data file and a plain-language question. The data goes to solvers gated against reference implementations before they were allowed to serve, never to hand-rolled code.
cross-check
Independent instruments read the same data. Agreement is certified; disagreement names the exact hole instead of blessing an answer.
return evidence
Every answer carries a contract: exact, tolerance-bounded, or heuristic status, diagnostics, and for optimization a bound or a completed proof.
the public record
certified against the hardest public tests
Certification claims are cheap; public benchmarks are not. The record here is independently published and checkable.
quplane's exact QUBO work holds first place on Hans Mittelmann's published Nonconvex QUBO-QPLIB benchmark as of 20 August 2026: fastest shifted geometric mean, 22 of 23 instances solved globally, and six instances that no other listed solver closes within the one-hour limit, one of them in 25 seconds against seven full-hour timeouts. Separately, the QPLIB library's best-known value for instance 2036 was improved here and verified by the library's maintainer. The proving ground holds the details, and proof of solves has planted-answer tests with downloadable datasets you can rerun.
faq
questions people ask
Can an AI agent return provably correct answers?
Yes, for a real class of problems. Optimization can carry completed optimality proofs, filters and estimators carry tolerance-bounded contracts, cross-checked instruments certify agreement. The honest boundary matters as much as the proof: where certification is not available, the system says so with a named status instead of a confident guess.
Is this the same as an agent harness?
It is the part most harnesses are missing. A harness decides what an agent may touch and what evidence it must return; most stop at permissions and take the agent's computation on faith. Certified computation is the verification member done structurally. The agent harness page covers the whole frame.
What does the certificate actually contain?
A typed contract beside the native result: status, tolerances, diagnostics, and for optimization problems an optimality bound or a completed proof. The contracts page shows the real format, all nine classes.
How does an agent call it?
POST a data file and a plain-language question. Every quplane instance
documents itself to agents at /llms.txt, /agents.md,
and /openapi.json on its own address; an agent pointed at the
instance learns the API without a human in the loop. It ships as an AMI on
AWS
Marketplace and runs entirely inside your own AWS account; no data ever
reaches the vendor.