Why we built Nyx with rules first

When you watch the public conversation about AI safety, almost all of it is downstream of one architectural choice: train first, constrain second. The model learns the world from a broad corpus. The constraints come later — RLHF, prompt engineering, output filters, content classifiers.

We started Nyx with the opposite premise. Define the behaviour first. The agent’s responses come from explicit JSON intents — text patterns paired with response templates — that you can read, audit, and edit. Machine learning extends that scaffold: intent classification, slot extraction, fluency. It does not replace it.

This isn’t a fashionable position. Rule-based systems went out of style in the late 90s for good reason — the long tail of “what could a user possibly say?” is enormous, and authoring intents for all of it doesn’t scale. We’ve felt that pain authoring Nyx. The rule-explosion is real.

But three properties of rule-first systems compensate, and they compound:

Auditability. A rule-based agent’s behaviour is explicit. You can read what it will do. You cannot do that with a 70-billion-parameter language model — not in any meaningful sense.

Update locality. Fixing a bad behaviour in a rule-based agent is editing a JSON file. Fixing the same bad behaviour in a trained model is a retraining cycle. The cost difference is several orders of magnitude.

Resource economics. Nyx ships in megabytes and runs on a laptop. A trained model that delivers comparable behaviour for a constrained domain costs tens of thousands of dollars to train and serves from a GPU. For the Foundation’s small-team economics, that’s not a side concern — it’s the whole concern.

The rule-explosion problem doesn’t go away. Our working hypothesis is that narrow ML components — small intent classifiers, small slot extractors — should sit under the rule scaffold, not in place of it. The rules remain the spec. The ML softens the edges.

This is what Rule-Before-Train AI Safety means in practice. Read the research initiative for the longer argument. Read the Nyx product page for what ships today.

The eval suite is on the roadmap. When it exists, anyone can argue against this position with comparable benchmarks. That’s the point.