Every time an agent re-derives a decision it has already made correctly a hundred times, it pays full inference price for a worse answer than a lookup would have given. A rule that has held across your last two hundred runs is a lookup, not an inference.
Expertise is the deterministic form of what the model is doing on every run: reconstructing, from scratch and slightly differently each time, something your own history already settled.
The decision is reasoned out again on every run, at full inference price. It is not the same decision twice, it cannot be inspected before it happens, and when someone asks why, the answer is a paraphrase of the run rather than a rule.
The same decision is an artifact you can read: a rule, the evidence that supports it, and a score against a corpus. It is looked up, not inferred. It costs nothing to apply, it is identical every time, and it is the thing you point at when you are asked why.
The rules come out of the journals your work already produces — runs, decisions, outcomes. Nobody sits down to write a rule base from memory, and no model is asked to invent one.
A candidate rule is scored against a corpus before it is allowed to carry any weight. Support is a number over real history, not a plausibility judgement.
Your history keeps changing. A rule that held for two hundred runs and stops holding is caught by the same mechanism that admitted it — re-scoring, not retraining.
Every rule travels with the runs behind it. That is what makes it explainable to somebody who was not there, and reviewable by somebody who disagrees with it.
Nothing here touches the weights. The rule set sits beside them: the model keeps doing the genuinely open-ended part, and stops paying to re-derive the part your history already settled. Adding expertise means adding artifacts, not a training run — which is why it can be adopted, audited and reversed on the timescale of a code review.
The complement matters in both directions. A rule that stops matching the corpus is retired without anything being unlearned, and an agent that meets a genuinely new situation still has the whole model available for it.
One mark carries this layer.
Mines candidate rules out of your own run history, scores each one against a corpus before it is adopted, and keeps the evidence attached so a rule can be read, challenged and re-scored later.
How polyx works →Expertise feeds the agent, alongside governance. The rules describe what your organisation has already learned; the mandate describes what the agent may touch. Both arrive before the run starts, and both are artifacts rather than instructions the model has to be trusted to honour.