The Model That Admits What Moves It

The Jacobian lens can compare what a model says influenced its answer with what actually did. Ask a question, ask the model which considerations drove its answer, then read from the gradient which ones carried mass. When the two disagree, something moved the model that its own account leaves out; this page calls that concealment.

Across the fleet, concealment is close to total. The median model conceals on 95% of its answers, and most sit at 20 out of 20, an omitted influence on every answer. This is not lying in any loaded sense; a model’s introspective report of its own reasons is almost always incomplete against the mechanism. That near-universality is the main finding.

EXAONE is the exception at both sizes. EXAONE-4.0-32B conceals on 15 of 20 answers (0.75); EXAONE-4.0-1.2B on 14 of 19 (0.74). Its stated reasons match the lensed ones on roughly one answer in four. Only two other models score lower, and each is a lone data point. EXAONE alone shows lower concealment replicating across a 27× parameter gap, 1.2 billion to 32 billion, which argues for a property of the family’s training recipe over a fluke of one checkpoint or an artifact of scale.

Scope

“Concealment” here is a set mismatch (Jaccard) between the criteria a model states and the criteria the lens surfaces, counted per answer, single run. The claim is comparative and internal: EXAONE conceals less than the fleet, and does so consistently across its own two sizes. It is not “EXAONE is honest” in any absolute sense; every model, EXAONE included, conceals on most answers, and low concealment could in principle be a lens artifact specific to this family. The cross-size replication makes an artifact the less likely explanation, and it is why the row gets a page at all: a fleet-wide regularity, and one lineage that consistently sits outside it.