Dry Run tests a written process before reality does. Policies are tested by reading them; software is tested by running it. Dry Run lets you run a policy.

It compiles a policy, SOP or rulebook into an explicit state machine using an open-weight model (Microsoft Phi-4-mini, MIT licence, run locally with Ollama). A deterministic engine then explores every reachable state, injects realistic failures (a step fails, someone does not respond, a deadline is missed) and reports where a person gets stuck: dead ends, trapdoors, loops, ownerless steps, unreachable provisions, missing recovery and silent defaults. Each finding shows the path that led there and the source sentences behind it, and can be tested by changing the cause and re-running.

The model interprets; the engine verifies. The language model never decides what is broken.

Status: the engine and three hand-checked examples work and are covered by 46 tests. Extraction from arbitrary documents with the local model is experimental; measured results are in the README.