Introducing the Research Collider: Goal-Conditioned Concept Collision for Grounded, Falsifiable Research Directions

Ask a large language model to connect two fields and it will, fluently — increasingly even aimed at a research question you pose. The hard part is no longer producing an idea but trusting it: a model shows its sources but not the reasoning behind the leap, and sounding sure is no sign of being right. We introduce the Research Collider, which wraps the Concept Collider (Wahl, 2026a) — two distant concepts colliding along a structural fracture — in a glass-box scaffold of selection and external certification, conditioned on a research goal: it steers a collision toward the goal’s answer-form, forges the result into a model that runs, and grounds its predictions against the literature. It thus occupies a gap no surveyed system fills — proposing and grounding a falsifiable research direction before it is carried out, not scoring finished papers or surfacing links the literature already holds. Shown end to end on modelling human forgetting, three findings frame it honestly: steering is a real lever on a direction’s form, not its fit; a free model told to diversify reaches the same real-data fit, so the value is an auditable process plus external certification, not better ideas; and the model’s own score is no guide to fit — on a second, fit-discriminating problem (the solar cycle) we find no evidence it tracks fit. The leads it returns are not guaranteed true, but guaranteed traceable, testable and aimed — which is what separates a research instrument from a novelty engine.Keywords: goal-conditioned creativity, scientific hypothesis generation, concept collision, white-box AI, falsifiability, research-direction generation