Response to EU Commission Draft Guidelines on High-Risk AI System Classification

1. About TeNova and SWT3

Tenable Nova LLC is a US-based company that develops the SWT3 AI Witness Protocol, an open cryptographic protocol for AI system accountability. SWT3 produces per-inference witness anchors that record compliance-relevant factors without intercepting, storing, or transmitting prompt content or model outputs.

The protocol is implemented as open-source SDKs (Apache 2.0 license) in Python, TypeScript, Rust, C#, and Ruby, published on PyPI, npm, crates.io, NuGet, and RubyGems. The current version (v0.5.3) covers 49 AI procedures across 22 namespaces, validated by 928 cross-language tests. The protocol specification is patent pending.

2. General Comments on the Draft Guidelines

TeNova supports the Commission's risk-based approach to AI classification and offers the following observations relevant to the guidelines:

2.1 Continuous evidence vs. point-in-time assessment

The draft guidelines appropriately emphasize that high-risk classification triggers ongoing obligations under Articles 9, 11, 12, and 14. We note that current conformity assessment practices rely primarily on point-in-time documentation reviews. For AI systems that evolve through fine-tuning, adapter stacking, RAG context updates, and multi-agent orchestration, point-in-time snapshots may not capture the full operational risk profile.

Cryptographic witness protocols like SWT3 provide continuous, tamper-evident evidence that bridges the gap between assessments. Each inference produces a fingerprinted record that auditors can verify independently at any future point.

2.2 Classification should consider evidence mechanisms

We recommend that the guidelines acknowledge the availability of automated evidence mechanisms when determining whether an AI system's risk can be adequately managed. Systems that produce continuous compliance evidence may present a different risk profile than systems with equivalent capabilities but no continuous monitoring.

2.3 Proportionate data protection through clearing levels

The draft guidelines raise valid concerns about data protection in the context of AI system monitoring. The SWT3 clearing protocol addresses this directly through four levels of data clearing (Level 0 Analytics through Level 3 Classified), allowing evidence fidelity to be proportional to data sensitivity. At Level 3, only numeric factors and hashed identifiers cross the network boundary.

3. Annex III Category Mapping

The following table maps each Annex III high-risk category to the SWT3 procedures that provide relevant continuous evidence.

4. Article 9: Risk Management System

5. Article 11: Technical Documentation

6. Article 12: Record-Keeping

7. Article 14: Human Oversight

8. Recommendations

TeNova respectfully recommends the following additions to the final guidelines:

1. Recognize cryptographic witness protocols as valid evidence mechanisms. Per-inference witness anchors provide a higher assurance level than periodic documentation reviews for demonstrating continuous compliance with Articles 9, 11, 12, and 14. The guidelines should acknowledge that automated, tamper-evident evidence mechanisms exist and can supplement traditional assessment practices.
2. Accept clearing levels as proportionate data protection. The guidelines should recognize that compliance monitoring does not require access to raw prompt content or model outputs. Clearing protocols that transmit only cryptographic hashes and numeric factors can provide sufficient evidence for conformity assessment while fully respecting data protection requirements under the GDPR.
3. Encourage interoperable evidence formats. We recommend that the guidelines reference NIST OSCAL and open witness protocols as interoperable evidence formats that Notified Bodies can consume without vendor-specific tooling. SWT3 anchors are verifiable with a single SHA-256 hash comparison, requiring no proprietary software.

9. Company Details