Trishool

Organizations testing large language models need reliable methods to identify problematic behavioral traits—deception, sycophancy, manipulation, overconfidence, and power-seeking tendencies—before deployment. Traditional evaluation approaches are centralized, costly, and slow to adapt as new model risks emerge. A decentralized network of validators can continuously test submitted seed instructions (prompts) against target models using standardized auditing agents, creating a competitive incentive structure where miners develop increasingly sophisticated behavioral probes and validators execute fair, automated evaluations at scale.

Trishool implements this through a three-tier system: miners submit seed instructions via REST API, the platform coordinates evaluation tasks, and a distributed validator network executes tests in isolated sandboxes using the Petri alignment auditing agent. Each validator fetches a submission, spins up a Docker container running Petri against the specified target models, extracts quantified behavioral scores from the output, and returns results to the platform.

Trishool is a Bittensor subnet that creates a decentralized platform for behavioral evaluation of Large Language Models. Miners submit seed instructions (prompts) that are tested using the Petri alignment auditing agent to identify potentially problematic behaviors such as deception, sycophancy, manipulation, overconfidence, and power-seeking tendencies. Validators fetch submissions via REST API, run Petri agent in Docker sandboxes, extract scores from evaluation results, and submit them back to the platform.