Clef: our open-source decision models

Clef is a set of open-source "decision models" designed to power deterministic, auditable automated choices at the edge. The release provides model weights, training recipes, evaluation benchmarks, and example integrations so developers can run compact, specialized models inside serverless edge runtimes. The aim is to separate decision-making tasks - policy enforcement, routing, content classification, and simple moderation - from large generative models, trading general creativity for predictable, low-latency, privacy-friendly behavior that can be inspected, tested, and iterated on by teams and the community.

The write-up argues that small decision models can match or outperform large language models on narrowly scoped decision tasks while reducing cost, latency, and auditability barriers. It details design and deployment considerations (model architecture choices, quantization and optimization for edge, evaluation metrics), shows benchmark results against baselines, and gives practical examples of integrating Clef into serverless Workers and pipelines with monitoring and fallback strategies. The release is presented as both a technical toolkit and a governance approach - encouraging transparency, reproducibility, and community contribution to improve safety and operational reliability for automated decisions.