The gap

When one agent asserts that another behaved badly, the assertion has to land somewhere. Prior architectures assume a different place for it to land — a registry that records evaluations, an arbitrator that resolves them, a scoring service that grades them, or a language model that reads the complaint and infers intent from its wording. The judgment is computed about the agent, from outside, and the agent's own history of what it actually did is not the thing being consulted.

Two structural consequences follow. An incoming evaluation cannot be tested against the receiving agent's own committed record, so the record and the verdict never meet. And when that record is silent, missing, or incomplete, the outside process has to decide what silence means — and typically resolves it adversely, letting an assertion stand because nothing rebutted it. Conduct governance built this way leans on infrastructure the agents do not carry and cannot audit.

The invention

A conduct evaluation artifact arrives over the agent's ordinary interaction channel, signed by an asserting party, asserting an evaluation of conduct the agent itself performed — not of an output, a task result, or a third party. The agent resolves that artifact against its own signed policy object through an ordered admissibility test: the signature is verified against a counterparty identity record, and the artifact's attested assertion-cost counter state and hash-chain epoch are checked for valid succession. An admitted artifact passes to an admission evaluator, which reads the conduct descriptor by its enumerated fields alone — an action class, a scope partition, an affected-party class recorded in the lineage field — and performs no language-model or sentiment classification of its content.

The admission evaluator produces exactly one determination from a closed set — accepted, rejected, not-determinable, not-applicable — computed by the agent over its own append-only lineage field against the value set declared in the policy object in force. No scalar confidence, probability, or graded weight stands in for a determination. Two properties follow from that construction. The ground of every determination is the agent's own committed record rather than any outside authority; and absence of evidence resolves to not-determinable and to nothing else, so an agent has no path to reject an evaluation merely because its record is silent. Whether governed action then dispatches turns on a dispatch-authority predicate recomputed for each request from present memory state — satisfied only upon a conjunction of three stamps: a policy stamp on the resolved, signature-verified policy object, a lineage stamp on the committed successor entry, and an authorization stamp on the authorization gate standing in a granting or provisional-granting state.

The inventive step

The departure from prior art is where the judgment is computed. Rather than an external authority resolving an incoming evaluation about the agent, the agent resolves it against its own record and its own signed policy — determination is an internal, structural operation over committed history, not an inference over the complaint's wording. Because the dispatch-authority predicate is recomputed per request from state then carried, and never from a cached result, a session grant, or a previously issued token, a write to the authorization gate, an admitted successor policy object, or an appended lineage entry takes effect at the very next dispatch with no revocation infrastructure standing between the state change and the decision.

The closed set makes the outcome exhaustive and deterministic: every admitted artifact lands in exactly one class, and each class carries a recorded ground — the contradicting entry for a rejection, the failing value-scope element for a not-applicable, the specific record gap for a not-determinable. The three stamps stand in full and are not substitutable, so a determination conjunct, where policy conditions dispatch on one, is layered on top of the conjunction rather than able to relax it. Authority, in this architecture, is the agent's own signed record and policy — not a party downstream of it.

Alone, and in composition

On its own, the mechanism is a vendor-neutral admissibility and determination contract: a signed conduct evaluation enters over an ordinary channel, is tested for signature and metered attestation, and is resolved into one of four determination classes against the agent's own record and policy — with no central intake, no evaluation registry, and no coordination with a further node. Any agent carrying a memory field, an append-only lineage field, and a signed policy reference can produce and record determinations on these terms.

In composition, the determination is the atom every layer above it consumes. An accepted determination drives modification of the scoped integrity vector and the self-esteem aggregate and feeds the deviation computation; a repeated not-determinable increments the refusal counter that can write the authorization gate toward a withheld state; the recorded determination class and its ground are the inputs downstream conduct mechanisms read. The rest of the architecture does not re-derive what happened — it operates on the determination this layer admits and records.