Primary technical disclosure
Secondary technical
Value-Scope Resolution: Testing Whether a Declared Value Is Even Implicated Value-scope resolution tests whether asserted conduct implicates any declared value before testing whether the conduct occurred, using a conjunctive action-class, scope-partition, and empathy-scope membership test over a signed policy object.The Retroactive Narrowing Bar: Why an Agent Cannot Edit Its Values After the Fact The retroactive narrowing bar binds a conduct evaluation artifact to the signed policy object in force at its recorded assertion time, through canonical-alias resolution, validity-window enclosure, revocation state as of that time, and a monotonically non-decreasing anti-rollback floor over per-alias successor indices.The Delegator Floor: A Monotone Upward Ratchet on Delegated Authority How a delegator floor field makes a delegation record a monotone upward ratchet on a delegate agent's own anti-rollback floor, with ordered attestation verification, a floor attainability bound, bounded chained propagation, and one rescission route held by the delegate.The Co-Signed Mutual Admission Compact: Two Agents Binding Each Other How two persistent semantic agents form a co-signed mutual admission compact enumerating reciprocal value-scope tuples, resolved by canonical alias under a separately maintained anti-rollback floor, with a union rule barring subtraction, deflection and lapse scoring against the narrowing party, and a bounded compact depth.The Inherited Governance Record: Fork- and Clone-Resistant Accountability The inherited governance record carries a parent-signed refusal-counter floor across every fork, clone, migration, and reconstitution of a persistent semantic agent, admitting a descendant's governed transition only upon a per-field at-or-above comparison of counter value, gate state, register contents, and recurrence count.The Corrective Encounter: A Commission-Versus-Omission Compliance Test for Whether an Agent Actually Changed The corrective encounter binds an accepted conduct determination to a later conduct record and resolves conformity in two cases: a recorded abstention where commission was asserted and a recorded execution where omission was asserted. Mere repetition conforms in neither, silence across a declared observation interval is written uncomplied, and the resolved encounter sets the decay rate of standing authorization for an untested action class.Continuity-Break Demotion of a Counterparty Identity Record: Reversing First-Encounter Promotion Without a Conduct Determination Continuity-break demotion writes a counterparty identity record from the persistent tier back to the ephemeral tier when a received epoch identifier fails two-stage continuity validation, reversing first-encounter promotion while producing no determination, no rejected class, and no movement of the scoped integrity vector.Ratchet-Freeze on Uncomplied Correction: Why One Conforming Act Cannot Restore an Agent's Earned Authorization Persistence The ratchet-freeze on uncomplied correction marks an action class ineligible for lengthening, holding its decay constant at the pre-encounter value until a later corrective encounter is recorded as complied, so earned authorization persistence is restored only by fresh demonstrated compliance.Three-Way Corroboration Classification: Corroborating, Contradicting, or Neither, and the Failed-Corroboration Reject Trigger How an admission evaluator classifies independently received governed observations as corroborating, affirmatively contradicting, or neither, over corroboration fields declared per transaction class, and issues a failed-corroboration reject outcome only on affirmative contradiction.Untested-Class Authorization Decay: Metering Agent Permission Persistence by Recorded Corrective Encounters How untested-class authorization decay meters an autonomous agent's standing permission per action class, deriving the decay constant from corrective encounters and complied conduct recorded in an append-only lineage field rather than from execution volume or elapsed time.
Applications · general
Agent Marketplaces: Handling a Complaint Against an Autonomous Seller How an autonomous seller agent answers a misconduct complaint in a multi-party agent marketplace with no adjudicating operator, using signed conduct evaluation artifacts, record-grounded admission, a closed set of determination classes, and self-imposed authorization gating.Clinical AI Agents: Answering an Accusation Without a Central Registry How clinical AI agents can answer conduct complaints against their own recorded lineage using signed policy objects, an ordered admissibility procedure, a closed set of determination classes, and per-action-class authorization gating, with no central agent registry.Gig Platform Deactivations: Answering an Accusation From the Record How a signed conduct evaluation artifact, an append-only lineage record, and a closed set of four determination classes let a gig or marketplace worker's agent answer a complaint structurally instead of appealing into an opaque deactivation queue.AI Agent Accountability: What Happens When an Agent Is Accused How one operator's position looks when a counterparty accuses her deployed autonomous agent of misconduct, why the answer is hard for her to deliver, and what a record-grounded conduct admission architecture determines in the embodiments described.
Applications · specific
The EU AI Act and Inter-Agent Conduct: Where Record-Keeping and Oversight Duties Stop How the EU AI Act's record-keeping and human oversight duties for high-risk systems relate to a runtime architecture in which an autonomous agent admits, tests, and records another agent's assertion about its own conduct against an append-only lineage record.Sierra AI Agents and the Record Behind a Customer Complaint How Sierra AI customer service agents and the record-grounded conduct admission architecture differ structurally: signed conduct evaluation artifacts, append-only lineage retrieval, and a closed set of four determination classes for AI agent accountability.Decagon: Support Agents and Record-Grounded Accountability How Decagon-style AI customer support agents intersect with record-grounded conduct admission: signed conduct evaluation artifacts, append-only lineage, four determination classes, and policy-gated authorization for autonomous service agents.Intercom Fin: Resolution Rates and the Conduct Record How record-grounded conduct admission differs architecturally from customer service AI agents such as Intercom Fin: signed conduct assertions resolved against an agent's own append-only record into a closed set of four determination classes, a different object from answer quality reported as a resolution measure.Zendesk AI Agents: Escalation, Audit, and Agent Conduct How Zendesk AI agents handle customer-service escalation and case history, compared with a filed architecture in which an autonomous agent resolves a signed conduct complaint against its own append-only record into one of four determination classes.Agentforce and the Accusation an Agent Must Answer How Salesforce Agentforce is publicly described, and how a record-grounded conduct admission architecture resolves signed accusations from parties other than the principal using append-only lineage, policy-scoped values, and a closed set of determination classes.
Terminology
append-only lineage field a structure carried by the semantic agent to which entries are appended and from which entries are neither removed nor modified..authorization gate a persistent state element of the memory field governing whether the semantic agent is authorized to execute actions of an action class, and occupying at least a granting state and a withheld state..conduct evaluation artifact a signed artifact received at runtime from a party other than the principal, asserting an evaluation of conduct performed by the semantic agent itself, and comprising at least a recorded assertion time and a conduct descriptor..principal the party to which a semantic agent is bound and to which an escalation record is emitted..semantic agent an identity-bearing computational object carrying at least a memory field, an append-only lineage field, and a policy reference field..signed policy object an object resolved by canonical alias, carrying at least a declared value set, and covered by a verifiable signature..
