ReproLoop

ReproLoop is an evidence-first bug-replay workbench for Android developers. A bug video shows a symptom; a regression artifact should also show what was done, what should remain invariant, and how an opposing control behaves. For one self-owned Cancel-persistence fixture, recorded Meta Muse Spark 1.3 inference connects video/frame context to a bounded semantic tap trace and fresh-screen action grounding. An independent app-state/event oracle checks the outcome. The same export is replayed with no model calls on faulty and developer-supplied fixed builds. The fixed build is a control with AI assistance, not autonomous source repair. The workbench presents the recorded source, saved post-action frames, faulty/fixed controls, and an exportable review packet. It does not perform a fresh device run or model inference. Imported claims remain UNVERIFIED; backend scores remain null. Missing evidence is not a passing result. The original pilot was 2/5, below its target. A separate sampled-context final cohort was 3/3, one attempt per case. These are different conditions and are not pooled. Both cover one bug family and three UI variants. Success required uniform frames, pointer pixels, temporal OCR, and supplied fixture/label context, not unaided general video understanding. A matched developer baseline also completed 3/3; no superiority or human-effort saving is claimed. The three-minute silent English-captioned demo is recorded, not live. Project-owned source is MIT licensed; third-party notices and AI-assistance disclosure remain separate. Independent-host Android and real novice validation remain incomplete unless separately evidenced.