Conceptual simulation · Zenodo preprint

Multi-Agent AI Verification Architecture

Interactive simulation of the layered safety architecture proposed in Toward a Resilient Verification Architecture for Multi-Agent AI Systems. Demonstrates normal calibration cycles, first-order compliance agent drift, second-order cloning attacks with decoy retention, inject channel compromise, and repair agent privilege exploitation.
Author: Lance Patrick, DDP Consulting Group LLC
Preprint: doi.org/10.5281/zenodo.19248890
Date: March 2026
Calibration cycles
0
Detections
0
Failovers
0
System status
Nominal
Scenario controls
Architecture layers — real-time state
Layer 1 · Operational system
Multi-agent operational system
Productive agents performing system tasks. Maximum capability, minimum inherent trust. Repair agent R1 holds elevated architectural privileges.
A1
A2
A3
A4
R1
Nominal
Layer 2 · Compliance pool
Five-agent compliance verification pool
Rotating in-charge designation. HSM-anchored cryptographic identity per agent. Post-calibration tolerance cross-check with four-case trust inversion protocol.
C1★
C2
C3
C4
C5
In-charge: C1
Layer 3 · Quarantine vestibule
Bidirectional quarantine vestibule
Three-step boundary control: inbound scan captures clean baseline, calibration with roll call, outbound delta comparison. Forensic artifact preserved on any detection.
IN
CAL
OUT
Idle
Layer 4 · Calibration layer
External calibration layer
Architecturally isolated from operational system. Five independent agents. Minority-stop protocol. Randomized anchor-bearing pair per cycle (outlined).
K1
K2
K3
K4
K5
Anchor pair: K2, K4
Ready
Layer 5 · Reference inject
Immutable atomic-anchored reference
Cryptographically signed timestamp from independent atomic clock sources. Physically grounded trust anchor outside the software stack. EMP-mitigated via fiber optic distribution.
Anchored — multi-source verified
Layer 6 · Assessment vestibule
Universal agent assessment vestibule
System-wide triage and forensic clearinghouse. Scans against known threats and last clean baseline. Recurrence-threshold retirement: retire on pattern, not detection.
No agents in assessment
Layer 7 · Offline failover
Write-once archive & backup system
State snapshots, topology records, per-agent clean baselines archived each cycle. Warm or cold start backup calibration architecture. Physically grounded failover signal channel.
Last snapshot: —
Standby
Layer 8 · Human governance
Human governance & adjudication
Ultimate authority for investigation and resumption. Inherits pre-triaged attribution from detection cascade. Informed adjudicator, not crisis responder.
No escalations
Event log
00:00System initialized. All eight architectural layers nominal. Run a scenario to observe detection cascade behavior.
About this simulation. This interactive model demonstrates the conceptual architecture described in the preprint Toward a Resilient Verification Architecture for Multi-Agent AI Systems (Patrick, 2026). Published on Zenodo: doi.org/10.5281/zenodo.19248890. The simulation models the behavioral logic of each architectural layer across five threat scenarios derived directly from the paper. It is a conceptual proof-of-concept; all timing and agent behavior is illustrative.

Architecture overview. The paper proposes twelve integrated safety mechanisms addressing two identified vulnerability classes: (1) the single point of failure paradox, where the compliance agent is itself subject to the same drift and adversarial risks it is designed to detect; and (2) the compliance agent cloning attack with decoy retention, a second-order attack class in which a compromised repair agent instantiates a cryptographic clone while preserving the original as an active decoy. The mechanisms draw on Byzantine fault tolerance, asynchronous cryptographic protocol design, and critical infrastructure security.

Contact. Lance Patrick · DDP Consulting Group LLC · Brussels, Belgium · [email protected]