Abstract

Abstract (Part 22 of 27)

Two independently administered AI agents agree to schedule equipment maintenance. The first organization’s policy allows inspection and coordination within an EU-hosted processing location. The second organization’s policy allows scheduling only for a registered task, denies external payment authority, and requires an auditable controller. A third-party supervisor proposes a joint workflow. Authentication proves who presented credentials, but it does not establish that both policies permit the same operation, that the system has discovered every affected resource owner, that the words “schedule,” “export,” or “completion” have equivalent semantics, or that the accepting runtime still enforces the policies at the moment of use. A signed policy may still be inaccurate, outdated, compromised, inapplicable to the selected resource, or incompatible with another valid signed policy.

Difference from earlier parts: Earlier embodiments control the joint plan, the freeze, evaluation, next-actor selection, role reconfiguration, selective communication, the next-move passport, Finality Sink enforcement, substitute selection, pending-work escrow, verified phase advancement, actual-outcome feedback, independent evaluator integrity, cross-domain partition recovery, protected session closure, cross-organizational supervisor appointment, and independent evaluator appointment. This embodiment specifically controls cross-domain trust and policy compatibility — determining whether a set of mutually independent policies and trust assertions can legitimately apply to the exact shared task at all.

This embodiment (Part 22 of a 27-part series under Indian Patent Application Number 202631131650; full details at https://zenodo.org/records/23940766) admits a cross-organizational task through a protected trust and policy compatibility procedure that derives affected authority principals from authenticated resource dependencies rather than from the requester’s declaration. Per-domain policy assertions are verified for authority, scope, schema, semantic meaning, generation, freshness, and any applicable delegation restrictions. A compatibility controller computes permissible shared activity, applicable obligations, and conservative upper limits without broadening an individual owner’s permission. Unsatisfied hard constraints, conflicting mandatory controls, ambiguous translations, unidentified owners, missing policy, or incomplete authenticated approval cause DENY, HOLD, or bounded renegotiation, not an implied joint grant. A canonical joint authority capsule binds the plan, participating authorities, policy digests and generations, restricted operations, supervisor identity, permitted sites, evidence requirements, duration, and use limits. Separate organizational approval and local installation evidence is required before a protected coordination interface becomes usable. Subsequent requests are checked against the same capsule, current protected policy, and resource ownership. The present embodiment does not equate secure transport, token possession, organizational membership, or matching text with actual permission.

The architecture therefore allows independent organizations to cooperate only on the exact subset of activity that all their policies actually permit, without treating authentication or similar wording as joint authorization.

Keywords: multi-agent systems, agentic AI, AI safety, AI governance, cross-domain trust, policy compatibility, semantic policy verification, bounded joint authority, fail-closed systems, zero-trust agents, joint authority capsule, non-delegable restrictions, execution finality, protected local admission, constraint intersection, policy generation fencing, independently authenticated policy, continuous validation, ambiguous translation rejection, resource-dependency discovery.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License

Download

Share

COinS