Semantic Conversational Intent Assist is the doctrine that Lucia understands lightweight utility, social, and scoped context prompts by meaning instead of exact phrase. The separate intent-assist model call that once carried it is retired from the served Focus route; the doctrine now lives in the one composition.
Dated: the intent-assist stage left the served route on 2026-09-04. Focus
answers from one brain (ALGO-40) retired the intent router and the
focus_ops_intent_assist model call from /admin/operator-focus. Fable reads
the whole house and composes the answer in one call; a greeting, a clock
question or a scoped property question is answered from the same pack. The
families below are preserved as doctrine and as the Focus bank’s must-not-say
and must-say expectations, not as a served classification stage. Guest-facing
intent tasks (guest_intent, guest_entity_extraction) remain separate,
bounded GPT calls.
Definition
Semantic Conversational Intent Assist was, at the Engine source verified on 2026-08-12 (9a15b1da199923d1ecb0927be2507feda2cb4c71), an Engine-owned Focus classification contract. That is a dated source statement. Since 2026-09-04 the same purpose is served by the one composition on Fable, whose pack carries the property clock and the last turns, so the small prompts below are answered in context without a router.
It classifies prompt families that are too small, social, or context-light for full operational triage but still matter to the human using Lucia.
It supports:
Evidence boundary
Use Focus Ops Intelligence for the served path, Intent System for the intent doctrine, and Lucia Intent Layer v1 for the dated historical implementation checkpoint.
Why Phrase Patching Is Rejected
Phrase patching fixes one string while missing the neighboring meaning.
It creates brittle behavior:
The Canon rejects phrase patching. Lucia must classify the semantic family, understand the human purpose, and preserve the product boundary.
Conceptual Families
Semantic Conversational Intent Assist protects families such as:
These families are not open-ended chat categories. They exist only to help Lucia respond naturally inside Focus with Lucia. On the one-brain path they are answered from the pack: the property clock answers “Time?”, the last turns answer “Let’s do this”, the stay in window answers a scoped property question.
Protected Prompt Families
Representative prompt examples in the contract include:
Lucia should not reject these as off-topic merely because they are short. She should also not inflate them into broad dashboard narration. Their inclusion documents the intended semantic families; it does not assert a fresh pass result for each exact phrase.
Guardrail
Lucia must not become open-domain ChatGPT.
The sublayer may help Lucia answer, acknowledge, clarify, or route a conversational prompt, but it does not expand Lucia’s domain beyond operator guidance, truth-state discipline, and the stable Focus Ops response contract.
Core doctrine:
See Also
