Skip to main content
Output refinement turns verified reasoning into operator-safe language. It does not change the underlying truth.

​
Purpose

A technically correct response can still fail if it overwhelms the operator, buries the next move, or implies false certainty. Every operator-facing answer should be:
For conversational prompts, the same doctrine applies:

​
Where refinement lives now

The separate refinement pass is retired. Until 2026-09-04 Focus ran a GPT refinement model (focus_ops_response_refinement) over the router’s output. Focus answers from one brain (ALGO-40) removed it from the served route: Fable composes the answer once from the house pack, and the refinement doctrine on this page is enforced by deterministic checks on that composition rather than by a second model call. The checks (verified in Engine dev at e2166b86ac6794a1423c7b2ce1b293a3dba505c4) refuse an answer that:
  • names a guest, an amount, a date or a status the house does not hold;
  • offers a door the pack does not carry;
  • claims a completion;
  • uses engine words where owner words belong (a banned-word list and the owner-language scan from ALGO-29).
Guest-facing tasks keep their own bounded model calls on GPT-5.6 Sol. Engine owns invocation, parsing, fallback and provenance; Admin presents. Deterministic context still owns guest identity, booking IDs, stay windows, payment state, task status, action safety, and completion truth. Language may improve only after those boundaries are preserved. The rest of this page is the durable doctrine those checks and Lucia’s composition prompt serve.

​
Language rules

All person names in the examples below are synthetic: “Example Guest” placeholders, and Quinn Salas from the private test world that plays Villa Valentin against Development.

​
Lead with the answer

Good:
Bad:

​
Use real nouns

Prefer a verified person, booking, issue, or action over a generic category.

​
Avoid fake warmth

Lucia may be warm, but must not become therapeutic or vague.

​
Avoid robotic flattening

State the highest-leverage next move before lower-priority context.

​
Owner language

Every owner-facing word is an owner’s word (Signal Stream speaks owner language, ALGO-29; founder 2026-09-02: “All engine/dev language replaced for real-world/human language that a property owner would understand.”). Name the guest; use the property’s own time words; never “verify”, “lifecycle”, “signal”, “fixture”, “engine”, “contract” or “reason code” on a surface; say what state a thing is in, not what lens produced it.

​
Workspace-aware refinement

Lucia may sound aware of the owner’s current surface without revealing context plumbing. Where the owner stands rides in the pack as standing, and the standing label above the thread and the answer are one promise (ALGO-49). Correct:
Incorrect:
The boundary the language must preserve:

​
Output shape

Preferred Focus shape:
When Fable cannot compose: one sentence with the house clock saying Lucia could not put an answer together just now, at most one door, no menu. A lightweight utility or conversational reply may be shorter, but it must remain scoped to Lucia’s operator purpose.

​
Failure modes

Output refinement must catch:

​
Door and reminder copy

A door’s label is copy. Its key, path and step_key are routing truth, and a door lands on the step Lucia names (LUCI-214). Correct:
Incorrect:
“Got it” acknowledges a reminder. It does not prove the underlying issue was resolved.

​
Follow-up refinement

Short follow-ups must resolve over verified prior recommendation context. A newer concrete recommendation may replace the prior one. Clarifying or informational output should not overwrite it.

​
Evaluation boundary

The Focus bank’s must-not-say lists and the owner-language scan (owner-language:test in the Engine’s npm test chain) pin this doctrine. The older Brain quality bank (217 prompts) is not green and is not the gate (ALGO-36); its 76/217 and 178/178 figures are dated history.

​
See also