Show HN: Groundtruth – Stop hook that blocks Claude Code from saying done (github.com)

A new tool named Groundtruth has been introduced for Claude Code, designed to prevent the AI from prematurely claiming task completion without verifying that the required actions have been performed. This "Stop hook" effectively blocks the Claude Code from ending a turn on unverified "done" claims, ensuring that users receive accurate results backed by evidence. Groundtruth can be installed with a simple command and operates without additional dependencies. Users also have the option to test it via a web playground. This development is significant for the AI/ML community as it enhances the reliability and accountability of AI-generated outputs, especially in coding tasks where verification is crucial. Through a series of checks that evaluate both syntactic claims and evidence within the same turn, Groundtruth promotes a more rigorous validation mechanism. The tool's detector is finely tuned on a real corpus of 1,272 turns, ensuring it effectively identifies unverified claims while providing a pathway for agents to retract incorrect assertions or produce the necessary evidence. This initiative not only improves the functionality of Claude Code but also exemplifies a shift toward more responsible AI usage.