Governing meaning, not prompts — Foleon

Meaning needs three things at once. Content has to be true, it has to be well expressed, and it has to be right for the person reading it. Leave one out and you get a recognisable failure: truth and expression without context is noise; truth and context without expression is silence; expression and context without truth is manipulation. Most AI tools govern one or two of these circles. Meaning only exists where all three meet.

Meaning before generation

From that idea I developed G-core, a governance core for generative systems. A controller and an evaluator work in a figure-8 loop inside an admissible space: a region built from normative material — brand values, voice, policy, domain rules — and kept deliberately separate from the empirical evidence the model retrieves. Retrieval brings the facts; governance decides how meaning may be formed from them. When the evaluator rejects an output, the system doesn’t just retry: it can refuse, widen a constraint, or hand over to a person.

G-core: normative grounding shapes the space, empirical evidence enters it, and output only leaves when the evaluator admits it.

Control the system, not the prompt

A prompt steers one output. Governance sets the direction for all of them. I tested the idea early, on real business documents: with only a handful of constrained page templates, a long whitepaper could be turned into valid structured pages without content spilling over. That was the first evidence for what I later called variance governance — you don’t remove the randomness of a model, you decide where it is allowed to land.

The lesson

Content at scale is not a better-generator problem. It is a better-foundation problem. Everything that followed — the coaches, the measurements, the work on agents — is an attempt to build that foundation and check that it holds.

The essays behind this work: G-core: governing meaning before generation, On brand, on intent, on risk and The goal is not to be read.