What the Model Does Not Say

Discovery is mostly the work of defining the problem: choosing the question and stating what is known. We take this literally. In a declarative engine, a gap is whatever the declared model fails to say, and a discovery is either finding such a gap or revising the question or the declarations so that it closes. We build on SEMPL, a Prolog engine that writes every process step as one fixed-arity semantic fact and admits no change that has not been verified and approved. Prolog suits this view structurally: it solves every problem with one algorithm, so the only lever on what it finds is the goal posed over the declarations. We pose discovery questions as goals. Counting the models of 38 declared processes found a contradiction between two authored facts and measured how far the engine's own hand-written repair departs from its declarations. Abduction returns minimal explanations that agree with an independent custody query on all 886 trials. An audit of every negation finds 127 places where "not declared" is silently read as "false" about the domain. Three symbolic routes then find missing structure, each a different kind of absence: chain of custody (seeded gaps recovered at 0.985–1.000 recall, 1.000 precision), outcome symmetry (652 material failure outcomes absorbed into 20 accounting processes, each verified) and analogy across sibling processes, which we show can only confirm what the declared outcome space already says. The same gate revises the question, the method and the model: the right one of two contradictory rules is kept on 3 of 3 seeds, a wrong question is told from a wrong method with precision 1.0, and one revision of the semantic layer cut the findings only analogy could make from 49 to 8. The general methods also coped with situations nobody had trialled them on. Without being seeded or anticipated, they found a reviewer approving their own trial balance, a physical-custody violation in a biological transport process, an authorization gap shared by three independently authored vocabularies (schema.org, PROV-O, W3C ORG), and, turned on the engine itself, three defects in its own governance, which were then fixed under the gate. Coping with the untried is what most definitions of intelligence name; we report it as generalization, not emergence. What did not hold is reported with equal weight: a pre-registered believability test failed, the repair breaks semantic custody rules in 16–37% of checks, absorbing outcomes enlarges the fragility surface, a new semantic class misfiles a quarter of what it draws, and none of seven pre-registered emergence indicators fired. Each failure points to more semantics, added as declarations, not to a different algorithm.