Skip to main content
This page preserves durable intent-layer doctrine alongside a historical implementation snapshot imported into the Library on 2026-06-23. The 47/47 result below is evidence for that checkpoint only, not current Engine identity or current evaluation health. See Current System State for current runtime, deployment, and evidence boundaries.
Snapshot status: Historical implementation evidence Checkpoint: hybrid intent assist layer and weather context utility validation snapshot Checkpoint validation: 47/47 strict eval pass

​
Purpose

The Lucia Intent Layer is the first explicit runtime layer responsible for helping Lucia understand what the operator means, not merely what phrase the operator typed. This matters because Lucia is not a generic chatbot and not a dashboard wrapper. Lucia must combine:
  • Operational intelligence — route the operator toward the right work.
  • Emotional containment — reduce pressure, preserve calm, and avoid making the operator feel managed by a machine.
  • Truth-state discipline — never claim facts, weather, actions, or completion states beyond what the system actually knows.
The intent layer exists to make Lucia more forgiving, more natural, and more intelligent without weakening operational boundaries.

​
Core doctrine

Lucia should understand the user’s goal before deciding the response posture. The correct loop is:
user languagedeterministic routingsemantic assist when ambiguousconfidence / arbitrationsafe routeoperational, human, or clarification response
The wrong loop is:
hard-code every phrase the user might sayhope enough phrases were anticipatedfall back to generic capability copy when not
Phrase lists may exist as scaffolding, but they are not Lucia’s mind.

​
What v1 established

At this checkpoint, Focus Ops documented a hybrid intent-assist layer with the following behavior:
  1. Deterministic routing stayed primary
    • Strong known cases routed without model interpretation.
    • Hard off-role boundaries still won.
    • Known operational categories continued to use explicit routing.
  2. Semantic assist for ambiguous cases
    • Model-assisted classification was limited to weak deterministic matches or unresolved follow-up context.
    • The assist result was structured, bounded, and reviewed by an arbiter before it could affect behavior.
  3. Confidence-gated arbitration
    • Assist was accepted only when confidence was high enough for the class of decision.
    • Low-confidence assist did not override deterministic behavior.
  4. Conversation context
    • Follow-up, recent social interaction, and fresh-request context could shape interpretation without erasing safety boundaries.
  5. Bounded weather-context utility
    • Weather could be operational because it affects arrivals, concierge work, outdoor plans, maintenance, and guest readiness.
    • Lucia clarified whether the user meant their current location or a managed property.
    • The checkpoint did not authorize invented or unsourced forecast data.
  6. Safe social language handling
    • Greetings, check-ins, gratitude, bounded humor, and valedictions were treated as valid human utility interactions.
    • Benign human language did not have to be treated as misuse.

​
Public behavior contract

The durable contract is about observable behavior, not a particular implementation layout.
  • Deterministic safety rules are evaluated before semantic assistance.
  • Semantic assistance is bounded to classification and clarification; it does not become an independent action authority.
  • Confidence and context may affect routing, but they cannot weaken hard role, identity, source, or action boundaries.
  • A final response must remain explainable in terms of the detected intent, relevant context, applied boundary, and available evidence.
  • Current implementation ownership is asserted only in Current System State and its dated evidence boundary.

​
Deterministic strength model

Lucia assigns deterministic confidence before asking for semantic help.

​
Strong

Examples:
  • hard off-role prompts
  • clear human utility
  • tiny factual
  • explicit operational lanes like payment risk, arrivals, maintenance, concierge
Strong deterministic routing should usually not be overridden.

​
Medium

Examples:
  • broad general focus prompts
  • weaker but still operational signals
Medium cases may not need assist unless follow-up state is unresolved.

​
Weak

Examples:
  • generic off-role fallback
  • ambiguous follow-up language
  • short compressed launch phrases
  • semantic property-context questions
Weak cases are eligible for assist.

​
Semantic families

The durable semantic families are:
  • Follow-up acceptance — compressed language that selects or continues an already offered next step.
  • Operational request — focus, triage, arrivals, payment risk, concierge, maintenance, defer-safe work, or overwhelm support.
  • Property-context utility — a bounded question whose answer depends on a property, guest stay, arrival, or external condition such as weather.
  • Human utility — greetings, gratitude, check-ins, acknowledgements, bounded humor, and closings.
  • Off-role — requests outside Lucia’s operational and human-utility role.
  • Unclear — language that lacks enough identity, object, property, time, or action context to route safely.
These families describe behavior. They do not prescribe a private wire format or implementation schema.

​
Arbitration principles

Semantic assist is not allowed to freely steer Lucia. The arbiter decides whether assist output can affect routing.

​
Hard boundaries

If the prompt matches a hard true off-role category, that boundary remains intact. Examples:
  • sports news
  • recipes
  • stock market
  • unrelated creative requests

​
Follow-up acceptance

If a pending follow-up exists and the user responds with compressed intent like:
  • “Let’s do this”
  • “All right, let’s begin”
  • “Start me there”
  • “Take me into it”
Lucia may use semantic assist to interpret the response as follow-up acceptance.

​
Social recent / soft-entry

When a recent human/social moment exists, Lucia may treat natural launch language as a soft entry into operations, but only within the allowed window and confidence rules.

​
Weather-context utility

Weather is not open-domain chat in Lucia. Weather is treated as a bounded property-context utility when the user asks about:
  • weather conditions
  • rain
  • outdoor plans
  • ceremony/dinner viability
  • guest readiness implications
  • arrival/concierge planning
Lucia should clarify scope before giving weather data.

​
Weather-context doctrine

Weather can be operationally relevant. Lucia should not answer weather requests as generic weather chat. Lucia should ask a bounded clarification such as:
or, when more than one managed property may be in scope:

​
Durable weather behavior

  • recognize weather and implied outdoor viability as property-context utility
  • clarify location/property scope
  • avoid generic off-topic fallback
  • preserve boundedness

​
Weather truth boundaries

  • inventing forecasts
  • providing weather data without a source/tool
  • implying integrated weather API execution
  • treating unrelated weather trivia as operational fact

​
Identity and guest-context doctrine

Intent classification must not invent identity. When language could refer to a guest, owner, operator, vendor, or another person, Lucia should:
  • preserve the distinction between the speaker and the person being discussed
  • clarify which guest, stay, property, or role is in scope before exposing protected context or taking action
  • avoid treating a name, message tone, booking reference, or conversational history as proof of identity
  • distinguish a request to explain, draft, verify, or act rather than collapsing them into one intent
  • keep uncertain identity or verification state explicit
Semantic confidence cannot substitute for authentication, booking verification, source evidence, or action authorization.

​
Human utility doctrine

Lucia should be able to handle safe human language naturally. Examples:
  • “Good morning”
  • “How are you?”
  • “Thanks Lucia”
  • “Nice to see you”
  • “Good night”
  • “Sleep well”
  • “See you tomorrow”
  • “Tell me a joke” within bounded hospitality context
Safe human language should not fall into generic capability fallback.

​
What this layer does not mean

This layer does not make Lucia a general-purpose chatbot. Lucia still should not drift into:
  • generic open-domain assistant behavior
  • unrestricted weather answers
  • sports/news/recipe/trivia mode
  • fake operational certainty
  • fake action completion
  • model-driven behavior without arbitration

​
Evidence standard

Intent work must leave enough evidence to explain a failure without publishing private diagnostics or operational controls. A review should be able to determine:
  • which semantic family was selected
  • whether deterministic routing was strong, ambiguous, or unavailable
  • which conversation context materially affected interpretation
  • whether semantic assistance was accepted or rejected
  • which identity, truth, role, or action boundary applied
  • what response posture ultimately reached the user
Threshold or routing changes require this evidence; wording alone is not a diagnosis.

​
Historical validation snapshot

On 2026-06-23, the recorded strict validation result was 47 of 47 passing scenarios. This checkpoint recorded the hybrid intent layer across:
  • social follow-up compression
  • valedictions
  • weather-context utility
  • hard off-role preservation
  • semantic property/weather prompts
  • end-to-end evaluation through the hybrid route

​
Historical evidence limits

The checkpoint had three important limits:
  1. Its weather expectations did not independently require both sides of the clarification: the user’s location and the managed-property context.
  2. Its weather behavior clarified scope only; the checkpoint did not prove a live weather source.
  3. A complete historical pass does not establish current model, route, session, deployment, or evaluation status.
Later claims about those surfaces require fresh evidence from Current System State or the owning system’s dated verification record.

​
Canon summary

At this historical checkpoint, Lucia Intent Layer v1 was the first working hybrid layer documented as interpreting ambiguous human language through a bounded semantic assist path while preserving deterministic safety, operational truth, and emotional containment. This is not the finished moat. It is the first real foundation of the moat.