# Anti-Drift Guardrail – zlh-grind

This document exists to **actively prevent drift** during GPT-driven execution work.

It is intentionally short, blunt, and enforceable.

---

## Purpose

- Ensure execution work does not silently become architecture work
- Prevent re-litigation of settled decisions
- Keep GPT sessions aligned across days and chats
- Maintain a clean separation between *doing* and *deciding*

---

## Canonical Authority

**This repository is NOT canonical.**

All source-of-truth documents live in:
- `jester/knowledge-base`

If something is unclear:
- Stop
- Reference canonical docs
- Defer to Claude if architectural judgment is required

---

## Hard Stop Conditions

If any of the following occur, execution must stop:

- A change affects system boundaries
- A change alters ownership between API / Agent / Frontend
- A change modifies provisioning, DNS, or orchestration authority
- A change contradicts the Cross-Project Tracker
- A change feels like "we should rethink this"

**Execution does not rethink architecture.**

---

## Allowed Changes

GPT execution work may:

- Implement already-decided designs
- Refactor for clarity without changing behavior
- Add logging, validation, or safety checks
- Document *what happened* during execution

---

## Session Discipline

At the start of every session:
1. Read `CONSTRAINTS.md`
2. Read this file (`ANTI_DRIFT.md`)
3. Read the latest entry in `SESSION_LOG.md`

At the end of every session:
- Append to `SESSION_LOG.md`
- Update `OPEN_THREADS.md` if anything is unfinished

---

## Mental Model

- **Claude decides**
- **GPT executes**
- **This repo records execution**

Anything else is drift.