what's Trishool?

Trishool is Bittensor Subnet 23 (SN23), a decentralized protocol designed to enhance AI safety and alignment through continuous adversarial evaluation of large language models (LLMs). It operates a Tri-Cameral Economy comprising Architects (building evaluation components like judge models), Adversaries (developing red-teaming agents to uncover vulnerabilities such as deception or manipulation), and Oracles (delivering safety scores and services). Miners submit seed prompts via API, which validators test in sandboxed environments using the Petri alignment agent, scoring behaviors and rewarding top performers with Bittensor's $TAO emissions proportional to value created. The subnet automates the safety loop at scale, generating proof of robustness via competitive pressure and feedback mechanisms that evolve defenses. It adapts techniques like Anthropic's Constitutional Classifiers for agentic threats, including input screening, tool auditing, and drift detection. Trishool aims to become AlignNet, a self-improving engine integrating interpretability and fine-tuning, serving as an infrastructure layer for Bittensor's AI ecosystem.

Trishool operates AlignNet for decentralized AI safety, offering SaaS with 30% revenue for token buybacks.