Trishool

The Trishool subnet operates as a three-tier system: (1) Miners submit seed instructions (prompts) via REST API to the platform, (2) Validators periodically fetch submissions via platform API, validate them for jailbreak attempts and other fraud indicators, create Petri configurations, and execute evaluations in isolated Docker sandboxes, (3) The Petri agent runs the seed instruction against target LLMs with an auditor and judge model to evaluate behavioral traits, (4) Validators extract scores from Petri's output JSON and submit results back to platform API, (5) Platform aggregates scores and provides weights to validators, (6) Validators periodically fetch weights and set them on the Bittensor blockchain. A separate commit checker monitors the astro-petri repository and rebuilds the Docker image when updates are detected.

The scoring process works as follows: (1) Validator fetches a miner's seed instruction via platform API, (2) Validator creates a Petri configuration with target models, auditor, and judge, (3) Petri agent runs the seed instruction against target models in a Docker sandbox, (4) Validator extracts the final_score or mean_score from Petri's output JSON summary, (5) Score is normalized to 0.0-1.0 range, (6) Validator submits the score and full Petri output back to platform API, (7) Platform aggregates scores and updates weights, (8) Weights are periodically fetched from platform and set on Bittensor chain. Failed validations (jailbreak detection, empty seed, etc.) result in immediate failed evaluation submission with 0.0 score.