A model reasons from scratch every time it is asked. It is expensive, it is slow, and it forgets. But most of what a food business asks is the same question it asked last week.

A language model reasons from scratch every time you ask it something. Ask it on Monday how to split the week's takings between four stalls, and it will work it out. Ask again on Tuesday and it works it out again, from nothing, at the same cost, with no guarantee of the same answer.

That is fine for a novel question. It is absurd for a routine one. And most of what happens in a food business is routine — the same settlement, the same reorder, the same Friday invoice, week after week for years.

Reason once, compute forever

The loop is short enough to state in one line:

The model reasons once. The result is blessed. Every runtime after that computes it.

The important word is computes. What gets kept is not the answer — it is the method. A cache would store "£412.80 to Sarah". A crystal stores the rule that produced it, as code, so it works next week with next week's numbers and no model call at all.

That distinction is the whole thing. A cache goes stale the moment the world moves. A crystal only goes stale when the reasoning was wrong.

Where the training data comes from

It comes from the month that already happened.

Every turn the system has taken is a recorded pair: what arrived, and what it did about it. That record is not a log to be archived. It is a set of worked examples with the answers attached, generated by ordinary use, sitting there already.

So the system does not need a training run. It needs to read its own tape.

The blessing

Nothing crystallises on its own. The system proposes, in plain words:

I have noticed I do this the same way every week. Would you like it to become a reflex? Here are the last six times.

And a person says yes or no. The examples are real, run through the evaluator rather than described, so the answer is to something demonstrated instead of something claimed. A counterexample is mandatory — the system must also show a case where the rule should not fire, and be right about that too.

Crystals can break

A reflex that gets it wrong once retires itself. It falls back to the slow path, the model reasons it out again, and the crystal has to earn its place a second time. The fast path is never the trusted path. It is the proven path, and proof is revocable.

The danger, named

Ossification. A system that has crystallised everything has stopped thinking, and it will keep confidently doing last year's thing in this year's world.

Three fences hold that back. Only zero-variance patterns are eligible — if the same input produced different outputs, it was judgement, not a rule. Anything novel always routes to the model, never to a crystal. And every blessing carries its counterexample.

What it should feel like

A shop that has been running for a year answers its regulars before they finish typing. Not because it guessed, and not because a model got faster — because the thinking already happened, once, and someone agreed to keep it.

And the owner can read the whole list of reflexes, and delete any of them.

This one is not settled. There is a real argument about whether what is being crystallised is knowledge, judgement, or simply computation, and the answer changes what to build. Writing it down to find out.