Skip to main content
The Intelligence Layer turns structured state, operator intent, and verified operational context into one calm, truthful next move.

What this layer does

Lucia’s intelligence layer exists to answer:
It does not exist to produce generic advice or broad dashboard summaries.

Verified current ownership

The current Development seam is split deliberately: The Development Engine at e2166b86ac6794a1423c7b2ce1b293a3dba505c4 (served at api-dev on 2026-09-10) runs two providers behind one gateway; the provider follows the model id. The configured default is gpt-5.6-sol. Fable (claude-fable-5-1) voices Focus: the Focus house composition and the payment-request draft are pinned to it on Development with no fallback (Fable is Lucia’s voice, ALGO-39, 2026-09-04; Focus answers from one brain, ALGO-40). The Fieldwork conversation brain and the guest-facing tasks stay on GPT-5.6 Sol. See Model Layer. The Algorithm that ranks the Signal Stream is deterministic code owned by one seat; the model is the voice and the hands, never the ranker. See LIEA doctrine. 0.1.3.6 is package metadata, not current release identity. Use commit SHA plus environment when identifying a runtime.

Primary intelligence surface

Focus with Lucia is the primary owner-facing intelligence surface, on the dashboard and in the docked panel beside every workspace step. Admin sends the request and where the owner is standing, and renders the response; Engine owns the pack, the composition and the provenance.
Since 2026-09-04 that is the only served path; the legacy intent router, intent-assist call and refinement pass are retired from the route (ALGO-40). See Focus Ops Intelligence.

Temporal spine doctrine

Calendar and booking state anchor intelligence to operational reality.
This is product doctrine. Individual routes and behaviors require current source or runtime evidence before they are called live; the admin surface words are documented on their own pages.

Workspace context doctrine

Workspace context may orient an answer, but it cannot override an explicit subject or change truth state. Lucia should support questions such as:
Operator-facing language must not expose raw request, context, transport, or response plumbing.

Guest-facing boundary

Guest-facing Lucia is a separate public concierge surface. It is not operator Focus Ops. Deterministic systems own:
Model-assisted language may express hospitality and ask for the smallest useful claim. It must not turn weak evidence into booking truth.

Current evaluation state

Three instruments, three questions:
  • The Algorithm replay harness (HLLC-29, Completed 2026-09-03) is the behaviour gate for the ranking: every Algorithm merge shows its old-to-new table, and “no behavioral change” is the explicit happy path.
  • The Focus bank (23 cases by 2026-09-07) pins Focus with Lucia: expected subjects, door keys and a must-not-say list, run in pack mode on every push and live on Fable at the gate.
  • The Brain quality eval bank (217 prompts) is not green and never was (174/217 and 192/217 on its two September runs; ALGO-36, Backlog). The standing gate for language changes is identical failing sets before and after, zero new failures. The older 76/217 and 178/178 figures are dated history.
Development and Staging are separate evidence domains. Staging serves Engine stage from 2026-04-28 and carries none of this.

Core responsibilities

Intent recognition

Identify the human purpose and preserve named entities, operational scope, and safe boundaries.

Operational ranking

Rank verified signals without replacing missing facts with model inference.

Emotional containment

Reduce cognitive load through fewer options, clearer action, calm tone, and truthful scope.

Output refinement

Lead with the specific next move. Do not bury it in generic prose.

The rule

Core doctrine:

Non-negotiables

Lucia must never:
  • overstate certainty
  • invent a system action
  • flatten urgency
  • bury the lead
  • turn configured defaults into provider proof
  • let another model speak in her voice when Fable cannot answer
  • give the operator more work than necessary

See also