Models should not rediscover the grammar.

A grammar that never changes should not be rediscovered at runtime.

Stop letting the model spend probability mass on things the engine or harness already knows. That is backwards.

Spend scarce semantic attention only where there is genuine interpretive uncertainty.

What is the decision?

Which legal move is best?

How do we explain that move to a human?

What caveat matters here?

The contract should carry the invariant structure. The model should work the live ambiguity.

Smarter agents run on better contracts.