A deterministic rule decides, in code you can read and test.
Risk-scored in seconds, so a human reviews the 5% that matters.
Powered By
Three properties that make the verdict something you can check, not just trust.
The notice-period comparison happens in plain Python, covered by 22 tests. A model that hallucinates cannot change the verdict.
A binary score removes reviewer ambiguity: 0 (safe) or 100 (critical). Anything the engine cannot read is returned as Undetermined, never guessed.
Granite's own verdict is parsed back and compared against the rule. If they disagree the rule wins — and the response says so, instead of hiding it.
Edit either box, or load one of the examples. This is the real rule engine from the backend repo, ported to JavaScript and running in your browser.
Risk score
A language model is good at reading a clause and unreliable at arithmetic. So it never decides the outcome — it explains one.
The last two are what a senior reviewer flagged at the hackathon, and they are the next things to build.
Recorded during the hackathon, November 2025: the agent running inside IBM watsonx Orchestrate and calling this project's audit skill. The panel above is the live version you can run yourself.