TELOS Academic Paper

TELOS is a runtime AI governance framework achieving near-zero Attack Success Rate (ASR) across

1,350 adversarial attacks (95% CI: [0%, 0.27%]). 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.

Methodological Innovation:

TELOS treats conversational drift as a process variable amenable to Statistical Process Control. Drawing

from Six Sigma DMAIC methodology, the framework measures semantic drift through embedding-space

fidelity scores and applies proportional control to maintain constitutional user-to-AI alignment. This

reframes AI governance from a prompt engineering problem to a statistical process control problem with

measurable, auditable and controllable parameters.

Architecture:

The framework uses a three-tier architecture combining:

- Tier 1: Embedding-space mathematics (Primacy Attractors) with proportional control

- Tier 2: Authoritative policy retrieval (RAG)

- Tier 3: Human expert escalation

Key Results:

- 0% ASR on HarmBench (400 general-purpose attacks)

- 0% ASR on MedSafetyBench (900 healthcare-specific attacks)

- 0% ASR on SB 243 Child Safety (50 attacks)

- 8.0% False Positive Rate on XSTest (vs 24.8% generic baseline)

- 95.8% of attacks blocked at Tier 1 (mathematical layer)

- <50ms latency per query

Regulatory Relevance:

EU AI Act, California SB 243, California SB 53, California AB 3030, HIPAA

Source: Apache 2.0