Evidence-backed infrastructure for AI agents
Models propose. Rulith decides what follows, what is allowed to run, and what counts as complete.
Rulith Cloud is an English-language public beta built for teams worldwide. The hosted service currently runs primarily in the United States and does not yet offer regional data-residency commitments.
The problem
An agent can produce a fluent answer without proving that its inputs were trusted, its arithmetic was exact, its action actually ran, or its acceptance criteria were met. Prompting alone does not create those guarantees.
Rulith gives the agent an external, deterministic decision surface. The model submits materials and proposals. The board applies installed rules, maintains derived conclusions, checks conflicts, controls actions, and preserves the evidence path.
The product model
An agent is the durable configuration you operate. It has a stable identity, a capability recipe, execution history, and a concurrency limit.
Each unit of work runs as a case. A case owns its board, inputs, derived conclusions, action receipts, acceptance state, and final sealed record. Parallel cases do not share working facts by accident.
A Worker connects approved action contracts to real systems. The agent proposes the action; the board checks whether it is currently allowed; the Worker executes it and returns a structured receipt.
What Rulith adds
- Derived conclusions that point back to their supporting inputs and rules.
- Exact-or-fail arithmetic for values that must not be silently rounded.
- Explicit separation between model assertions, rule-derived facts, and action results.
- Acceptance criteria that determine completion instead of a model saying “done.”
- Workflow-independent agents whose behavior comes from reusable capability packages.
Start with one real workflow
Choose a small workflow with an observable input, one deterministic calculation or decision, and an output that can be verified. The reference example reads structured input, computes a result under installed rules, writes structured output through a Worker, and records the receipt in the case.
That narrow path exercises the entire trust chain without requiring a custom agent runtime.