Abstract

Abstract (Part 8 of 27)

A neutral evaluator may detect an incorrect delivery destination, missing safety prerequisite, inconsistent currency, unacceptable jurisdiction, defective payment amount, or incomplete prior phase. If it simply tells an agent in natural language to “fix the plan,” the agent can alter an unapproved field, trigger an external API call, contact excluded peers, or silently modify the signed plan while other agents remain frozen. A later evaluator can then review a different candidate from the one the organizations originally authorized. A corrective instruction is not an unrestricted delegation, and an unapproved model-generated proposal must never overwrite the actual committed plan.

This embodiment (Part 8 of a 27-part series under Indian Patent Application Number 202631131650; full details at https://zenodo.org/records/23940766) implements the protected correction procedure. After the plan is committed, frozen, and evaluated, the independent evaluator issues a typed, non-effective correction direction that names the specific deficiency, target actor, fields, and policy context. A correction-turn issuer generates fresh single-use permission limited to the selected actor, permitted fields, allowed peer or message class, and current plan version. Protected gateways mediate any selective one-turn communication without reopening all channels. Proposed changes are written only into a separate candidate-plan staging register and tested against schema, materiality, and policy rules. The authoritative committed plan remains unchanged until all required participants re-approve the exact revised version, local owner vetoes are respected, and a new freeze converges on the updated plan. No evaluator directive, candidate delta, plan signature, or re-freeze certificate alone authorizes an external effect. Uncertain message or effect outcomes stay UNKNOWN and block unqualified continuation.

The architecture therefore allows controlled, auditable revision of multi-agent plans while ensuring that correction turns cannot silently become unrestricted negotiation or unilateral execution.

Keywords: multi-agent systems, agentic AI, AI safety, AI governance, protected plan revision, evaluator-directed correction, controlled renegotiation, non-effective corrective instructions, fail-closed systems, zero-trust agents, multiparty recommitment, single-use correction turns, candidate-plan staging, selective one-turn communication, cross-organizational collaboration, execution finality, versioned plan revision, owner-authorized correction scope, freeze re-convergence, non-effective plan deltas, protected multiparty renegotiation, field-scoped revision permissions, independently enforced plan correction.

Creative Commons License

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

Download

Share

COinS