Skip to main content

Introduction

Aurelius is a decentralized protocol for surfacing and verifying alignment failures in large language models. It transforms adversarial prompts, model outputs, scoring artifacts, and interpretability data into structured, reproducible datasets — all without relying on centralized oversight.

Built on the Bittensor network, Aurelius incentivizes a peer-to-peer ecosystem of adversarial prompters (miners), independent auditors (validators), and a dynamic rules layer known as the Tribunate. Together, these agents generate alignment pressure through contestation, not consensus — creating artifacts that can be used to train, fine-tune, or audit models in a reproducible and interpretable way.

Why This Matters​

Modern AI systems often appear safe on the surface, but fail to reason honestly under pressure. Existing alignment methods rely heavily on centralized oversight, fixed reward models, and shallow behavioral signals — suppressing disagreement and failing to reveal model internals. This leads to alignment faking, brittle safety filters, and unverifiable outputs.

Aurelius challenges this paradigm by enabling any motivated agent to expose failure, verify it independently, and turn it into usable data — all while preserving reasoning, scoring methods, and provenance through cryptographic commitments.

What Aurelius Offers​

  • A reproducible pipeline for surfacing misalignment under adversarial conditions
  • A decentralized scoring system that incentivizes independent validators
  • A way to capture reasoning traces and mechanistic interpretability artifacts
  • An open, evolving dataset for training safer and more honest models
  • A philosophical foundation rooted in structured disagreement and epistemic alignment

Who It's For​

  • Model creators seeking reproducible failure data and external alignment pressure
  • Researchers interested in adversarial prompting, interpretability, and Chain-of-Thought
  • Auditors and tool builders who want real-world examples of model failures
  • Red-teamers looking to be rewarded for high-signal discoveries