ICOSA: Multi-Model Byzantine Fault Tolerant AI Compliance Assessment

The EU AI Act (Regulation 2024/1689) introduces the world's most comprehensive regulatory framework for artificial intelligence, with enforcement beginning August 2, 2026 and penalties reaching EUR 35 million or 7% of global annual turnover. Despite this urgency, no scalable, objective, and verifiable compliance assessment methodology exists.

This paper introduces ICOSA (Independent Council for Objective System Assessment), a patent-pending protocol that applies Byzantine Fault Tolerant (BFT) consensus across a geographically and architecturally diverse council of AI models to produce compliance assessments that are objective, reproducible, and cryptographically attested.

By eliminating single-model bias, introducing multi-stakeholder deliberation, and sealing verdicts in a proprietary cryptographic Audit Chain, ICOSA creates the first trustworthy bridge between AI providers and regulatory requirements.

Current approaches to AI compliance assessment suffer from fundamental limitations:

• Single-point-of-failure bias: Assessments performed by a single model or auditor inherit the biases, blind spots, and limitations of that single perspective.
• Lack of reproducibility: Without standardized methodology, the same system can receive contradictory assessments from different evaluators.
• No verifiability: Traditional audit reports are static documents with no mechanism for independent verification or tamper detection.
• Scalability constraints: Manual auditing cannot keep pace with the volume of AI systems requiring assessment as obligations phase in through 2028.
• Regulatory ambiguity: The EU AI Act's requirements are complex and subject to interpretation, requiring multi-perspective analysis.

In illustrative scans against the EU AI Act, the most common gaps cluster around transparency obligations (Article 13), data governance (Article 10), and human oversight provisions (Article 14) — the problem is systemic, not confined to any single vendor or architecture.

ICOSA addresses these challenges through three interlocking innovations:

The ICOSA platform consists of five primary subsystems:

1. Assessment Engine: Receives system descriptions, contextual data, and observation framework inputs. Orchestrates the council deliberation process.
2. Council Manager: Maintains the roster of participating models, manages model health, diversity metrics, and rotation schedules.
3. Consensus Engine: Implements the BFT voting protocol, manages deliberation rounds, and computes final verdicts.
4. Audit Chain: Signs and appends verdict hashes to ICOSA's proprietary hash-chained ledger for immutable attestation.
5. Observation Framework: A guided data collection system that structures the information needed for comprehensive assessment.

The architecture is designed for horizontal scalability, with each subsystem independently deployable and the council size configurable per assessment tier.

The ICOSA council is composed of 11 AI models selected for maximum diversity across four dimensions: provider, geographic region, model architecture, and assigned deliberation role.

Illustrative seat composition — describes the diversity design, not a roster of specific named third-party models or vendors.

ICOSA adapts practical Byzantine Fault Tolerance (pBFT) for AI compliance assessment. The protocol tolerates up to f faulty nodes in a council of n members, where:

For the three ICOSA tiers:

The consensus process proceeds in three phases: Pre-Prepare (assessment distribution), Prepare (independent model evaluation), and Commit (vote aggregation and verdict computation). Dissenting models must provide structured reasoning, which is preserved in the audit trail.

Every ICOSA certification verdict is sealed in ICOSA's proprietary cryptographic Audit Chain — not a public cryptocurrency blockchain. The attestation includes:

• SHA-256 hash of the complete verdict payload, chained to the previous record's hash
• Ed25519 signature over the record, and the rolling Merkle root at time of append
• Timestamp and sequence number for temporal verification
• Quorum achievement proof (number of agreeing models vs. required threshold)
• Reference to the regulation version assessed against

This creates an immutable, independently verifiable record that cannot be retroactively altered without breaking the hash chain. Regulators, auditors, and stakeholders can independently verify any ICOSA certification by recomputing the hash chain and checking the signature against ICOSA's published Audit Chain public key.

The ICOSA Scanner is the assessment input pipeline. It collects structured data about an AI system through the Observation Framework — a guided questionnaire that captures:

• System purpose, intended use, and deployment context
• Training data provenance and governance practices
• Risk management system documentation
• Human oversight mechanisms and escalation procedures
• Transparency measures and user-facing disclosures
• Technical documentation and record-keeping practices
• Accuracy, robustness, and cybersecurity measures

This structured input ensures that every council member evaluates the same comprehensive dataset, enabling meaningful comparison and consensus formation.

Illustrative pattern observed across common deployment architectures — not a completed audit of specific named systems:

The most common deficiencies observed were in transparency obligations (Article 13), data governance (Article 10), and human oversight provisions (Article 14). These patterns underscore the case for automated, scalable compliance assessment as regulatory obligations phase in.

ICOSA is positioned as a complementary tool for the emerging EU AI Act compliance ecosystem. The protocol does not replace human legal judgment but augments it with scalable, objective, multi-perspective analysis.

Key regulatory considerations:

• Article 9 (Risk Management): ICOSA assessments can serve as a component of an organization's broader risk management system.
• Article 11 (Technical Documentation): ICOSA reports provide structured documentation that satisfies transparency requirements.
• Article 43 (Conformity Assessment): While not a notified body, ICOSA provides preliminary conformity assessment that organizations can use to prepare for formal evaluation.

ICOSA is designed to operate within regulatory sandboxes (Article 57) and can be adapted for national AI strategies across EU member states.

10.1 Transatlantic Alignment: CISA, Five Eyes, and NIST CAISI

On 1 May 2026, CISA and the cyber agencies of Australia, Canada, New Zealand, and the United Kingdom — the full Five Eyes partnership — jointly published Careful Adoption of Agentic AI Services. The guidance defines five agentic-AI risk categories and explicitly requires cryptographically anchored agent identity with short-lived credentials. ICOSA's architecture independently satisfies these requirements at the protocol level, providing an operational foundation that aligns with the transatlantic regulatory direction:

• Privilege risk: Least-privilege access enforced by capability-token validation; each agent action carries a single-use, time-bounded credential.
• Design and configuration risk: Boundaries encoded in the deterministic rule engine rather than delegated to the agents themselves.
• Behavioral risk: BFT-consensus council detects misalignment through multi-model deliberation rather than single-model judgment.
• Structural risk: The Overwatch lattice provides visibility into cascade dynamics across interconnected agent networks.
• Accountability risk: Every action carries an Ed25519-signed capability token; the full authorization chain is reconstructible post-incident.

On 5 May 2026, the NIST Center for AI Standards and Innovation (CAISI) expanded its pre-deployment frontier-model testing agreements to Google DeepMind, Microsoft, and xAI, joining OpenAI and Anthropic. CAISI evaluations target cybersecurity, biosecurity, and chemical-weapons risks across more than forty completed assessments, some conducted in classified environments by the interagency TRAINS Taskforce. ICOSA is positioned as the complementary downstream layer: where CAISI evaluates foundation models before release, ICOSA evaluates deployed AI systems against the same risk taxonomy in production-use contexts.

The EU AI Act represents a watershed moment for AI governance. With obligations phasing in through 2028 and systemic gaps observed across common deployment architectures, the need for scalable, trustworthy compliance assessment is acute.

ICOSA addresses this need through a novel combination of multi-model deliberation, Byzantine Fault Tolerant consensus, and cryptographic Audit Chain attestation. The protocol eliminates single-model bias, ensures reproducibility, and creates verifiable compliance records that stakeholders across the ecosystem can trust.

As the regulatory landscape evolves, ICOSA's modular architecture enables rapid adaptation to new regulations, updated requirements, and expanded council membership. The protocol is designed not just for the EU AI Act, but as a foundation for global AI compliance infrastructure.

Patent Pending: App. No. 63/789,142
© 2026 KYMA Tech Solutions / Sage Holdings LLC