CausalPatch: Count-Constrained Causal Repair for User-Edited Multi-Shot Narratives

Abstract

Interactive AI short-film systems should support local storyboard revisions without breaking later events or regenerating an accepted narrative. We study minimal executable causal repair: given an 8–10-shot structured storyboard, world state, and typed intervention, a model predicts dependencies, affected shots, and a minimal patch over Keep, Rewrite, Delete, and Insert. CausalPatch combines predicted graph propagation, a causal copy gate, global edit-count supervision, count-constrained decoding, and node-wise realization. We construct Formal-v2.1, a 2,000-case executable benchmark with multi-hop effects and template-disjoint splits. Count-constrained planning improves three-seed replay from 0.728 to 0.857 on test and from 0.752 to 0.888 on a post-freeze confirmation split. Under oracle content, it outperforms a frozen DeepSeek-dsv4 reference (0.857 versus 0.555 replay) while regenerating fewer shots; with a shared frozen realizer, replay rises from 0.558 to 0.645. A 40-case naturalized stress test exposes the remaining transfer gap.