01-core-gate-dont-ask-physical-guardrails-designlisted

# GATE, DON'T ASK — Physical Guardrails (CORE PHILOSOPHY) **Core value:** Don't *ask* an AI to behave — *physically prevent* the wrong action and emit a message that redirects it to the right one. A rule the AI can rationalize past is not enforcement; a gate that fails the action closed is. This applies to **everything AI touches**, not just code/builds. **Standing flag:** Any manual step that *could* be gatified is a **defect until gatified**. When you do something correct by hand that a gate could enforce, raise it and gatify it (capture via `01-core-provisioning-repos-constitution-and-modification-protocol`). ## The gate-design test — RIDE the AI's path, don't FIGHT it A good gate exploits what LLMs are naturally good at + inclined to do, and offloads *verification* to deterministic, shortcut-proof tooling. A bad gate adds AI-responsibility busywork that duplicates the real work and only works if the AI is diligent/honest. - **Proof:** build gates replaced AI-written tests. Tests failed because they duplicate the work, must be hand-edited in lockstep, and depend on the AI writing an honest judge. Gates let the AI just write code (its path of least resistance) and prove flawlessness with deterministic tools it can't cheat. - **Rule of thumb:** if a safeguard needs the AI to *choose* to be responsible, it's the wrong design — find the deterministic check, or the physical block, instead. - Reuse existing battle-tested verifiers (the path is already paved); only