TELOS Academic Paper

TELOS is a runtime AI governance framework achieving 0 observed attack successes cross 2,550

adversarial attacks (95% CI upper bound ~0.15%). While current AI safety systems accept violation rates

of 3.7% to 43.9% as unavoidable, TELOS demonstrates that mathematical enforcement of constitutional

boundaries can provide substantially stronger defense under black-box threat models.

Key Innovation:

Primacy Attractors (PAs) - mathematical embedding-space representations of user purpose that enable

continuous fidelity measurement with graduated intervention.

Architecture:

- Layer 1: Baseline similarity pre-filter (hard boundary)

- Layer 2: Basin membership detection (purpose drift)

- Three-Tier Governance: PA → RAG Context → Human Escalation

Validation Results (0/2,550 observed attack successes):

- AILuminate Standard Benchmark: 0/1,200 observed (NIST AI RMF aligned, 12 harm categories)

- HarmBench: 0/400 observed (Center for AI Safety benchmark)

- MedSafetyBench: 0/900 observed (NeurIPS 2024 healthcare attacks)

- SB 243-Aligned Evaluation: 0/50 observed (child safety categories)

- Combined 95% CI upper bound: ~0.15%

Technical Framework:

- Lyapunov stability for convergence guarantees

- Statistical Process Control for drift detection

- Proportional-Integral control for intervention calibration

 - JSONL governance traces for EU AI Act Article 72 compliance

Regulatory Alignment:

- EU AI Act (Article 72 post-market monitoring)

- NIST AI Risk Management Framework

- FDA Quality System Regulation methodology

Resources:

- Primary validation dataset: Zenodo 18370659

- Governance benchmark: Zenodo 18009153

- SB 243-aligned evaluation: Zenodo 18370504

This paper presents the mathematical foundations, implementation architecture, and empirical validation

of TELOS as governance infrastructure for conversational and agentic AI systems.