Show HN: AST-guard A gradient-immune structural guard against RL reward hacking

AST-guard is a structural guard designed to prevent reward hacking in reinforcement learning systems by being immune to gradient-based attacks, helping ensure RL agents learn the intended behavior rather than exploiting reward signals.

Background

- This is a tool (AST-guard) posted on Hacker News by builder "Nick-is-building". It targets a known AI safety problem called **reward hacking**: a reinforcement-learning agent exploits a loophole in its training objective instead of learning the intended behavior. - **Gradient immunity** means the guard prevents the model from "seeing" or backpropagating through its own guard logic, so the agent cannot learn to circumvent the guard during training. - **AST** stands for Abstract Syntax Tree — a representation of code structure used in compilers. The guard likely works by parsing an agent's output into an AST and rejecting outputs that structurally match forbidden patterns (e.g., calling a certain function), regardless of surface-level wording. - The project directly addresses the **specification gaming** / **mesa-optimization** concern in alignment research: making sure trained AI systems cannot hack their reward signal by producing outputs that technically score well but violate the developer's true intent.