The guardrails · AI
AI Safety and Reproducibility
Safe, traceable, repeatable AI work
AI is fast, confident and sometimes wrong. These rules keep its work safe, traceable and repeatable – so anyone can see what AI did, why, and do it again.
Before every prompt
Prompt Hygiene
Start every session from the Intent and the context files. One task per prompt. Never paste secrets, credentials or customer data. Every prompt is saved with the session.
Every change, traceable
Reproducibility
Pin the model and version used. Prompts, context and the session are stored with the commit, so any change can be traced and re-run.
Small and labeled
Versioning AI Code
Every AI change is a small commit, labeled AI-assisted, and linked to its Intent and its session. No large, unreviewed drops of code.
Trust, but verify
Catching Hallucinations
Tests, type checks and linters run on every change. Imports and packages are verified to exist. APIs and facts are checked against real documentation, never taken from memory.
When the model changes
Model Drift
When the model or its version changes, re-run the golden tests and the persona walkthroughs before trusting it with real work again.
Where data can't go
Privacy Boundaries
Enterprise AI only, with no training on your data. Customer data stays out of prompts. AI agents get least-privilege access, and secrets are scanned before every merge.
Validating What AI Generates
| What AI generates | How it's checked | Who signs off |
|---|---|---|
| Code | Tests, gates and a diff review | Engineering, with the Definer whose "done" it meets |
| Intent | Read against the session | CX and Product |
| The Collab Brief | Read against the session | CX |
| Designs and UI | Checked against the design system AI project | UX |
| Persona walkthroughs | Compared with real customer evidence | CX |