Skip to main content
Lucia combines deterministic operational structure with bounded model-assisted language and judgment. The model does not own operational truth.

​
Core principle

Lucia should not rely on model output alone.

​
Current model layer

The Development Engine at e2166b86ac6794a1423c7b2ce1b293a3dba505c4 (served at api-dev, 2026-09-10):
Fable voices Focus and drafts the payment request (Fable is Lucia’s voice, ALGO-39, 2026-09-04; Lucia asks Quinn for a new card, LUCI-210). The Fieldwork conversation and the guest-facing tasks stay on GPT-5.6 Sol. See Model Layer. The package value 0.1.3.6 is metadata. It is not a release identity and does not prove which commit is deployed. The configured model is not proof that a model ran for a particular request; every answer’s meta.model_runtime.tasks[] records provider, resolved model, latency, fallback and usage, and that record is the authority.

​
Deterministic responsibilities

Deterministic code owns:

​
How Focus reasons now

Since 2026-09-04 Focus with Lucia reasons from one pack and one composition (Focus reads the whole house, LUCI-189; Focus answers from one brain, ALGO-40): the Engine builds the house-state pack deterministically, Fable composes the answer and its doors in one strict-JSON call, and deterministic checks refuse anything the pack does not hold. There is no intent router, no refinement pass and no fallback model on that path. When Fable cannot compose, Lucia says so in one owner-language sentence with the house clock and offers at most one door she already holds.

​
Model-assisted responsibilities

When invoked successfully, the model may assist with:
The model must never own:

​
Runtime provenance boundary

/admin/operator-focus carries runtime and model-execution evidence in meta while preserving the stable top-level ok / data / meta contract.
Interpret those fields per response:
A 200 response does not prove provider success. An attempted call does not prove a response arrived or parsed. Admin may render this evidence, but it does not originate, repair, or infer it.

​
Current evaluation evidence

The Focus bank (23 cases by 2026-09-07, pack_ok in pack mode and live on Fable at the gate) pins Focus; the Algorithm replay harness gates the ranking. The Brain quality eval bank (217 prompts) is not green and is not the gate (ALGO-36, Backlog); the 76/217 and 178/178 figures on older records are dated history. See LIEA doctrine.

​
Operational reasoning boundary

Verified operational context can orient the response, and an explicit prompt subject must win over page context. Saved workspace state can inform the next response without converting a saved step into a resolved issue. Payment reasoning is bounded by the ledger: the money story is told in obligation order from Stripe truth and owner-recorded rows, and a payment item closes only when the ledger says confirmed_paid. Lucia proposes a door (send a payment link, charge the card on file, record a payment taken another way); the owner’s tap executes it. See Payment truth — current. If Fable cannot compose, Lucia degrades honestly: one sentence, at most one door, nothing invented, no other model in her voice. Core ordering:

​
See also