Rules tier · in build
Structured Output Assurance holds an AI's output to a schema you supply: JSON Schema, XML, or any format expressible as a formal grammar. Conformance is enforced as the output is generated, so a non-conforming output is never produced, rather than produced and caught.
What it does
Acceptance is binary and decidable. There is no score, no threshold and no calibration to argue about: the accepted set is the policy, and an output either belongs to it or could not have been produced.
A grammar automaton masks invalid tokens in the decoding loop, so a non-conforming token cannot be sampled. The output conforms because no other output was reachable.
Where the decoding loop is not reachable, the full response is parsed, accepted, or repaired and re-validated. An output too convoluted to parse triggers a re-prompt, never an unchecked pass.
How it works
Model
Any model, rented or self-hosted, generates token by token.
Boundary
Your schema, compiled to an automaton, removes every token that would break it.
Output
What reaches your system can only be valid against the schema.
Who it is for
Regulated entities
A claims, credit or payments system that consumes AI output directly. A schema-valid response on every call removes a class of integration failure, and with the Conformance Pack leaves a record of it.
Operators
A maintenance or process system fed structured instructions by an AI. Every instruction arrives in the shape the control system expects, or not at all.
What it does not do
A well-formed output can carry the wrong content. Governing meaning is the semantic check in Attested Semantic Compare.
Many conformant outputs can still drift together. That is what the Aggregate Drift Service watches.
Where an insider or a capable adversary is in scope, the rung is the hardware appliance.
Evidence rung it fills
That an output met a declared policy.
Structured Output Assurance supplies the accept or reject decision. The Certifiable Conformance Pack adds the tamper-evident record per output that an auditor or supervisor can rely on.
Related
Same mechanism, for actions
The automaton pointed at a tool registry instead of a schema.
Bounded Command →The evidence layer
Identical verdicts on re-execution, and a record per output.
Conformance Pack →How we grade claims
What is demonstrated, built, designed or still a hypothesis.
Status legend →Get in touch
Thirty minutes on where your AI acts, what evidence it would need to leave, and whether this is the right rung.
Book a 30-minute conversation →