G-core decides how something may be said before the model looks up what to say. Most generative AI works the other way around: retrieve some facts, write a prompt, generate, and hope the result fits the brand. G-core — short for governance core — reverses that order. That one change is what makes the output governable.
A newsroom already works like this
A good newsroom does not let reporters decide the house rules story by story. The style book and the editorial standards exist before anyone picks up the phone: how we name people, what counts as a confirmed source, what we never print. Reporters then bring in the facts. An editor checks the piece against both before it runs.
Two kinds of material, kept apart on purpose:
- How things may be said — brand values, tone of voice, policies, the rules of a domain. I call this normative grounding.
- What is true right now — facts, figures, sources. That is empirical grounding.
Together, the normative rules mark out an admissible space: the region of everything this organisation could acceptably say. Generation happens inside it, or not at all.
Why retrieval alone can’t govern
Take a bank’s assistant answering a question about mortgage rates. Retrieval can fetch today’s rate perfectly. It cannot stop the model from phrasing that rate as a promise, dropping the required disclaimer, or sounding cheerful about a rejection. Those are not facts to look up. They are rules about how facts may be used.
That is why, in G-core, retrieval is an evidence supplier, not a governance mechanism. It brings what should be said. The normative layer decides how it may be said — and it is in place before the first fact arrives.
The model gets two narrower jobs
Instead of one opaque generation step, the model works in two bounded roles:
- Navigator — inside the admissible space, it decides what to retrieve and how to interpret it.
- Evaluator — it checks the composed answer against intent, policy and context, and escalates to a person when it can’t make it fit.
A rejected answer doesn’t simply get regenerated until something passes. The system can refuse, ask for a wider boundary, or hand over to someone who can decide.
What that buys you
- Change behaviour without retraining. Adjust a policy or a boundary, and every future answer moves with it.
- Explain every output. Each answer traces back to its sources (empirical) and the rules it was held to (normative).
- Learn in the right place. Feedback sharpens the rules and the knowledge, not a prompt someone has to remember to update.
G-core is not about making generation more impressive. It is about making every output explainable: true to its sources, faithful to its rules, and right for the moment it is used in. That is what moves generative AI from a creative tool to a system you can put your name under.