Decision model-based tool call guardrails for Claude Code, Codex, and Cursor

A decision-model-based guardrail tool for AI coding agents such as Claude Code, Codex, and Cursor now ships with one-line CLI support that blocks unwanted file deletion, file creation, and malicious attacks. Fitted and calibrated on a large corpus of synthetic data and evaluated on 46 held-out proposals, the guardrail's accuracy rose from 60.9 percent to 95.7 percent, with an expected calibration error (ECE) of 0.065. The author notes that an ECE close to zero indicates a more trustworthy result.

Decision engine as your agentic, super-fast guardrail or gate. We have added CLI support; with just one line, your full claude code/codex/cursor, etc., is protected from unwanted file deletion, creation, or malicious attacks. Fitted and calibrated on a large corpus of synthetic data, 46 held-out proposals. Before: accuracy 60.9 percent. After fit+calibrate: accuracy 95.7 percent, ECE at 0.065. ECE close to zero is a more trust worthy result. prompt in first comment