Show HN: Hush – GitHub issue triage that abstains when it isn't confident (github.com)

Hush, a novel GitHub Action, has been introduced to enhance issue triage by abstaining from labeling when the AI model lacks high confidence in its assessment. Unlike typical LLM-based bots that can generate misleading labels based on uncertain predictions, Hush implements a decision model called Jev, which provides a probability score alongside its classifications. This approach ensures that only when the confidence level exceeds predefined thresholds—like 80% certainty for categorizing a bug—does it apply a label. If it can't confidently identify an issue, Hush remains silent and includes a rationale for its decision. This innovation is significant for the AI/ML community as it addresses a common issue with automated issue triage: the tendency of AI to provide inaccurate classifications, leading to poor repository management. Hush's design emphasizes precision and transparency, allowing developers to fine-tune thresholds according to their specific needs. Additionally, its cost-effectiveness (approximately $0.00002 per issue) makes it feasible for widespread use across projects without overwhelming the system. By prioritizing accuracy over activity, Hush sets a new standard for AI tools in software maintenance, promoting a more reliable interaction between developers and automated processes.