Core concepts

Agent

An agent is a durable product identity and configuration. It carries a name, description, stable ID, capability recipe, credentials, and run history. The language model is replaceable; the agent's governed behavior is not stored in the prompt alone.

Case and board

A case is one bounded unit of work. Its board is the authoritative working state for that case.

The board contains materials, goals, rules, derived conclusions, action definitions, action results, conflicts, and completion state. A case can be sealed when its board-enforced completion conditions pass. Sealing preserves a read-only record.

Working memory and closure

When operations are submitted, Rulith validates the batch and recomputes the rule closure. A conclusion appears because a rule matched supported inputs—not because the model wrote the conclusion directly.

Truth maintenance removes derived conclusions when their support is no longer active. Archived facts remain part of history but no longer participate in the active closure.

Three origins

Rulith keeps three origins visibly separate:

• Asserted — proposed by a model or caller.
• Derived — produced by the rule closure from supporting facts.
• Effect — produced by applying an action definition.

An asserted or effect fact does not become derived merely because it has the same shape as a derived conclusion.

Evidence tiers

Evidence strength belongs to the supporting chain. A weak input cannot be made stronger by passing through a rule or by changing a field name. Higher assurance must come from an accepted verification source or authenticated execution receipt.

Actions and Workers

An action definition is logical: name, preconditions, consumed facts, and produced facts. An action request is only eligible when its parameters can be bound from supported board state and its Constitution gates pass.

A Worker is physical: it performs the external operation and reports the result. The board records both the approved request and the returned receipt.

Completion

Completion is a board decision. Goals have explicit acceptance tests, pending obligations must be cleared, and required actions must have acceptable results. A model can propose that work is complete, but cannot certify its own proposal.