Zero-Shot Ontology Alignment Without Tears: a Deterministic Pipeline, Its Limits, and What Structure Actually Buys You

We present a zero-shot ontology-alignment pipeline that requires no per-pair training and no per-pair tuning: OWL parsing with synonym resolution, deterministic lexical hashing, a frozen pretrained encoder, greedy 1:1 assignment, and a fixed threshold sweep. On OAEI anatomy it scores F1 = 0.900 (competitive with trained matchers, though below the Matcha frontier), with recall on non-trivial pairs (recall+) of 0.774 against Matcha's 0.82; a full 21-pair conference matrix under one frozen configuration demonstrates breadth, with an erratum-corrected combined mean of 0.666 on small references. We contribute an evaluation triad — recall+, coherence rate, calibration error — and two measured negative results: hierarchy propagation and functorial (containment) coherence repair both destroy more value than they create, because cross-ontology hierarchies violate containment nearly independently of mapping correctness. We frame alignment categorically (span, approximate functor) as analysis language, and outline the open program of functorial regularizers.