Scope-Blind Verification, and What a Learned System Cannot Represent

Systems that reason exactly over typed structures — energy-based reasoners, constraint solvers, proof-carrying search, industrial and legal knowledge graphs — derive their central advantage from checkability: a conclusion can be verified because the constraints are explicit. This paper advances two claims about what such verification does NOT establish. Both grow more dangerous as the verification machinery improves. Claim I — scope-blindness. The type system a reasoner operates over is not derived but chosen, and the choice encodes a jurisdictional and professional carving of the domain which every conclusion inherits. A derivation formally correct under one carving can be materially wrong under another, and because the derivation checks, nothing signals the conflict. We give an executable demonstration in which identical facts yield opposite conclusions about whether a statutory protection attaches, under two documented carvings of the same relation. Claim II — a representational limit, which we think is the deeper claim. Some standards are normative-ideal: they specify what a competent or diligent agent ought to do, which may be what few agents actually do. A system trained on observed conduct produces, by construction, a central tendency of that conduct. These are different objects, and the substitution of the second for the first is invisible in the output and undetected by verifying the derivation. The companion Cryptotype LLM Probe dataset measures the adjacent effect directly, finding that the models' unmarked "neutral" is in fact a marked, cross-lingually imposed value profile that flattens real cultural differences — replicating and extending the WEIRD-default literature across US, EU and Chinese-origin models, under a frozen pre-registration and with the continuation-fluency confound removed by matched minimal pairs. Law supplies the sharpest instance, because Roman law distinguished the two explicitly — culpa levis in abstracto, measured against the bonus et diligens pater familias, versus diligentia quam in suis rebus, measured against actual habitual conduct — but the question generalises to any domain in which a system is asked to apply a norm it learned from behaviour. We locate both claims in four fields (healthcare, law, industry, marketing), propose a remedy costing three metadata fields — origin, authority and jurisdictional scope recorded on the relation type, so that "verified" becomes "verified within scope S" — state precisely what would refute us, and set out four open problems, of which the last (whether an open standard can be represented as open) we regard as genuinely unsolved. This deposit includes the preprint and the executable demonstration (scope-blind-demo.py), which uses no inference engine: the point is a property of the representation, not of the reasoner. Circulated as a debate opener. Refutation is the preferred response.