THE SAFETY ARCHITECTURE

TOOL-LEVEL VALIDATION Tool-level validation means each tool checks its own inputs and conditions, regardless of who calls it. This is the principle of never trusting the caller. For example, a file-deletion tool refuses paths outside its allowed folder, checks that the path does not contain tricks like going up directories, caps the number of files, and refuses system files. A payment tool validates that the amount is positive and below a limit, and that the account exists. Tool-level validation catches mistakes from any source: a confused model, a malicious injected instruction, or even a bug in another part of the harness. It is the layer closest to the real-world effect, and therefore a very reliable last line of defense.