For production video systems — safety, compliance, robotics — a 92%-accurate nondeterministic model is unshippable. A model that gives a different answer to the same input on different days cannot be audited, tested, or trusted as an engineering component. A deterministic model with known failure modes, on the other hand, is something you can build a system around: you can write tests for it, version it, and reason about its behavior under load.
Darwin's perception runs in embedding space with calibrated confidence scores. Language enters only at the query layer — never the perception layer — so outputs are repeatable. Given the same video and the same prompt, Darwin returns the same verdict, the same confidence score, and the same timestamped evidence, every time. That's not a nice-to-have. For safety, compliance, and robotics applications, it's the minimum bar.
Related: The case for deterministic video AI.
VLM hallucination mechanics and reproducible side-by-side clips — coming during our August launch.