Abstract

Abstract (Part 3 of 27)

Keeping the Snapshot Stable: Ensuring the Plan an Evaluator Sees Is Still the Plan the Agents Hold

Even after agents from different organizations have agreed on a joint plan and a communication freeze has been installed, the snapshot an independent evaluator examines can still become invalid. A counter-offer already sent can remain in transit, a retrying worker can still hold an old credential, an agent’s local compiler can apply a later translation, a restored virtual machine can revive an obsolete epoch, or an intermediate cache can show different state to different participants. The result is a classic time-of-check/time-of-use error: the evaluator reviews Plan X while one or more agents are already operating against a different or incomplete version of the plan. A signed plan alone is not a stable evaluation snapshot, and a freeze certificate alone does not guarantee that the associated evidence remains complete and consistent.

This embodiment (Part 3 of a 27-part series; full details at https://zenodo.org/records/23940766) adds the integrity-preservation and continuity layer on top of the joint-plan commitment (Part 1) and the verified bilateral freeze (Part 2). It binds the exact canonical plan bytes and schema to a causal communication cut, the disposition of every pending message and deferred operation, and the protected state generation of each covered domain. A causal event registry records every governed message, acknowledgment, retry, cancellation, quarantine and worker delegation with its predecessors and outcome classification. A pending-work ledger keeps unresolved items explicit; an unconfirmed external effect is never treated as canceled simply because a local job descriptor was deleted. After the freeze is effective, each domain constructs a protected local snapshot that is locked to the accepted transcript cut, fence generation, policy generation and coverage map. A cross-domain reconciliation controller compares the snapshots, checks causal compatibility, detects equivocation or rollback, and withholds evaluation on material disagreement. The resulting group snapshot convergence object is the only reference the independent evaluator is permitted to assess. Any subsequent change that alters a material plan predicate, causal predecessor, participant identity or authority generation invalidates the snapshot and returns the system to a hold. Recovery after restart, migration or partition restores or rejects state against the protected generation before any continuation is considered. The evaluator’s directive remains non-effective; final authority stays with each domain’s own enforcement boundary.

The architecture therefore turns a signed plan and a freeze certificate into a continuously revalidated, causally complete evaluation reference that fails closed on uncertainty—closing the gap between what the evaluator checks and what the agents can still do in multi-agent and agentic AI systems.

Keywords: multi-agent systems, agentic AI, evaluation snapshot integrity, causal message containment, time-of-check time-of-use, fail-closed systems, zero-trust agents, cross-organizational collaboration, pending-work reconciliation, state continuity, snapshot validity, AI safety, AI governance, distributed agent coordination, protected negotiation state, independent evaluator, execution finality.

Creative Commons License

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

Download

Share

COinS