Get oriented.

A proposition the system holds to be true, with a confidence score, provenance, and history. Beliefs are the atoms of the API. Everything else operates on them.

The full graph of beliefs in a session at a moment in time. Versioned implicitly: each accepted belief, each confidence patch, each new piece of evidence shifts the state.

The container that holds a belief state. Scoped to your API key. Sessions are isolated. Different keys cannot read each other’s sessions even if IDs collide.

The collection of raw sources (documents, notes, URLs, structured records) you feed in. The corpus is what beliefs get extracted from. The belief state is what you reason with.

A contradiction between beliefs, or between a belief and new evidence. Dissonance is the central signal: it is what the API is built to surface, measure, and route to resolution.

The opposite of dissonance: evidence or beliefs that reinforce another belief. Used to compute confidence and binding strength.

A 0–1 score on each belief. Patched as new evidence arrives. Below a threshold, beliefs go dormant rather than disappearing.

How tightly a belief is connected to the rest of the graph. High binding = removing this belief reshapes a lot. Low binding = peripheral. Surfaced on graph and metrics endpoints.

A session must be activated before search and inference resolve at full quality. Activation primes embeddings, indices, and edges. Cheap, idempotent, fast.

Deriving new beliefs from existing ones plus incoming evidence. Async. The API returns a job handle; results stream back as new beliefs in the session.

/who pulls actors and entities. /what pulls claims and actions. Two complementary lenses on unstructured text. Both feed beliefs.

Agent-scoped state stored alongside beliefs but treated separately. Use it for ephemeral situational awareness that should influence retrieval without becoming a permanent belief.

The raw signal that supports or contradicts a belief. Returned by the dissonance endpoint when you ask “why is this flagged?”

A human (or agent) signal that a belief should be promoted from candidate to held. Locks confidence above the dormant threshold.

A typed relationship between two beliefs: supports, contradicts, depends on. Edges are how dissonance is computed.

A cluster of beliefs whose mutual contradictions sum to a dissonance score above a threshold. Useful for routing. Group-level resolution beats one-belief-at-a-time triage.