Give agents a controlled
execution path
Rulith sits between model output and external side effects. It keeps evidence, policy, execution, and completion in one inspectable case.
One loop from proposal to proof
1. Establish facts
The model may propose identifiers and goals. Sources retrieve authoritative data and attest what they actually observed.
2. Derive the next action
Knowledge rules calculate what follows. Missing evidence remains missing; it is never promoted by repetition or fluent wording.
3. Execute through a Worker
Typed action contracts bind parameters from board facts, enforce preconditions, and expose only approved credentials.
4. Close with a receipt
Outcomes return as structured evidence. A case completes only when its acceptance conditions are derived.
Keep the runtime generic
The Agent sees Actions. Stable Tool references connect them to local execution without exposing implementation details to the model.
Knowledge + Actions
Vocabulary, rules, and governed operations.
Sources
Evidence identity, scope, and trust.
Worker Tools
Versioned capabilities mapped to local Adapters.
Constitution
Independent prohibitions and approval boundaries.
Run the local JSON workflow
Download the workflow-independent REPL and Worker, process one input file, and inspect the evidence-backed result.
Start the 5-minute runInspect the result