C1
Shared approval admission
Test at-most-one successful authorization admission across heterogeneous executors under declared races, duplicates, crashes, recovery, and partitions.
Research / Multi-Agent Systems
Runtime safety for multi-agent AI systems
A bounded research program for shared authority, delegation, coordination under partial failure, and independently checkable effect evidence.
Research object
Three bounded research hypotheses
The unit of analysis is the complete multi-agent workflow, not an isolated agent or tool call. Each hypothesis can fail under the declared model; a counterexample rejects the corresponding claim.
C2
Approval–decision–effect safety
Test the coupled invariant under explicit failures, bypass attempts, evidence failure, clock uncertainty, and ambiguous destination outcomes.
C3
Omission-aware effect evidence
Independently check causal evidence while distinguishing attempted, accepted, committed, observed, and unknown effect states.
Claim discipline
Current product boundary
Quiet Open-Core v1.2.0 is the current public baseline: fail-closed interception, deterministic policy, bounded human approvals, sandboxing, and local hash-chained evidence. It provides an experimental foundation; it does not by itself prove the MAS hypotheses.
Broad novelty, exactly-once external effects, production-qualified multi-tenancy, and complete evidence are not claimed. Final claims depend on a source-by-source comparison, a declared failure model, specialist review, and independent evaluation.