SPECTER: Document-level Representation Learning using Citation-informed Transformers

Embedding-based science mapping places every paper in a shared semantic space, inviting a forecasting inference: two communities drifting together should foreshadow their fusion. We subject that inference to a pre-registered retrodiction test, treating known cross-field fusions in mathematics and mathematical physics as forecasting targets and asking whether an abstract-level proximity signal rises in the years before the catalytic paper — every success criterion, term set, null model, and verdict fixed in a locked pre-registration before any similarity is computed. Run naively, the instrument would have supported a positive headline; over a five-case campaign we found and killed several apparent precursors — a saturated above-null proximity that passes a level test trivially, an anti-proximal z ≈ −4 that measures set coherence rather than intellectual distance, and a z-score descent that is an artifact of a growing set scored against a shrinking null. The paper's primary contribution generalizes these into a taxonomy of nine artifact classes, each with its exposing case, its guard, and its observable signature, in a reusable guard cascade. On the two fusion events the cascade admits as fully scoreable — one import-type (three channels), one convergence-type, the latter guard-complete from birth and verdict-grade at null — abstract-embedding proximity shows no above-random prospective convergence signal, a null that a two-directionally-validated coherence control separates from the artifacts. We state the result at exactly its scope: on these two events, at yearly-to-semi-annual resolution, with one embedding model on abstract-level text, semantic proximity carries no prospective fusion precursor — not a general law that field fusions are semantically unforecastable. The durable output is the instrument, its artifact taxonomy, and a governance ledger in which the same pre-declared machinery corrects the pipeline, its human advisor, and its operator alike: a demonstration that a fully disciplined science-mapping pipeline can produce a trustworthy negative.