Understanding & the Context Layer

Atune turns behavior into answers in two stages. The separation is deliberate: the grounding is computed without an LLM, and the language model only explains what the grounding already contains.

Signals → Context Layer → answer

1. Signals. Every captured behavior is a signal — an observable fact about an account, user, or feature.
2. The Context Layer. Signals are continuously composed into a structured grounding bundle per entity: activity trends, engagement, relationships, risk indicators. This is computed deterministically, without an LLM.
3. The answer. understand() composes that bundle into a plain-language, cited answer. The model explains the grounding — it does not invent facts.

Grounded and cited, or honest

Every answer is grounded in that workspace's real signals and cited. The shape:

type UnderstandResult = {
  knows: boolean;                 // false ⇒ "not enough understanding yet"
  answer: string;                 // grounded in your signals
  confidence: number;             // 0..1
  citations: { signal: string; value: string; when: string }[];
  recommendedNext: string | null; // one action, only when confident
  signals: Record<string, unknown>; // the structured grounding bundle
};

The most important field is knows. When there isn't enough understanding yet, Atune returns knows: false — it does not guess. An honest "I don't know" is more useful than a confident fabrication, and it's a hard rule, not a tuning knob.

Why two stages

Keeping the grounding LLM-free makes answers reproducible and auditable: the signals bundle and citations show exactly what an answer rests on. You can trust a recommendation because you can see the facts underneath it.

Next: how identity makes answers read "Kari at Acme" — Identity, PII & the DPA.