MAIVA: Multi-Agent Integrity Verification Architecture

MAIVA is the fourth component of a four-part autonomous authority stack (SATA, HMAA, CARA, MAIVA). It extends HMAA's individual authority model to multi-agent swarms with Byzantine-resilient aggregation, CUSUM-augmented anomaly detection, graduated escalation, and DoDD 3000.09 action gate classification.

Publication

DOI: 10.5281/zenodo.19015517
Author: Burak Oktenli
Affiliation: Georgetown University, M.P.S. Applied Intelligence
License: MIT

Authority Stack

• SATA: Hardware trust (TPM-anchored sensor attestation, DOI: 10.5281/zenodo.18936251)
• HMAA: Individual authority computation (DOI: 10.5281/zenodo.18861653)
• CARA: Authority recovery after collapse (DOI: 10.5281/zenodo.18917790)
• MAIVA: Multi-agent swarm trust aggregation (DOI: 10.5281/zenodo.19015517)

Key Components

• HMAA individual authority: A = τ_eff · C · H · (1−D) with temporal decay
• SAAP Byzantine-resilient trimmed weighted median, per-partition independence
• 3-layer detection: static + CUSUM temporal + enrollment drift
• Action Gates: CRITICAL/ELEVATED/ROUTINE/PASSIVE with coalition cap
• Graduated escalation with per-level action permissions and override requirements
• Sensitivity analysis: authority and F1 vs Byzantine ratio and agent count
• WCET profiling against 16ms budget
• 37 self-tests (21 functional + 4 property-based + 4 detection + 4 checksum + 4 CUSUM)
• TLA+ formal specification (3 agents, 5 invariants)

Files

• maiva-complete.html: Full simulation + 14-tab PDR briefing (2,396 lines)
• MAIVAAgent.tla: TLA+ formal state machine specification
• MAIVAAgent.cfg: TLC model checker configuration
• CHANGELOG.md: Full version history (v5.0 through v5.18)
• MAIVA_Punch_List.md: Engineering review punch list

Author

Burak Oktenli
Georgetown University, M.P.S. Applied Intelligence
ORCID: 0009-0001-8573-1667