> ## Documentation Index
> Fetch the complete documentation index at: https://helloluciallc.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Reasoning Model

<Note>
  Lucia’s reasoning model combines deterministic operational structure with model-assisted language and judgment.
</Note>

***

## Core Principle

Lucia should not rely on model vibes alone.

The system must combine:

```text theme={"dark"}
structured state
deterministic ranking
verified context packs
workspace context
model refinement
stable output contract
```

***

## Current Model Layer

Lucia Engine v0.1.3.6 currently uses:

```text theme={"dark"}
gpt-5.6-sol as default (GPT-5.6 Sol)
OpenAI Responses API for Lucia JSON gateway calls
```

The upgrade path is intended to future-proof model swaps while preserving Lucia’s behavior contract.

***

## Deterministic vs Model Responsibilities

### Deterministic Logic Owns

```text theme={"dark"}
route shape
response contract
known records
guest identity
booking identifiers
stay windows
payment state
Stripe financial truth
durable ledger truth
task identifiers
task status
route/action safety
save/completion truth
environment routing
deduplication
```

### Model Reasoning Owns

```text theme={"dark"}
language refinement
natural tone
some prioritization expression
multi-turn interpretation
emotional containment
operator-facing framing
permission-based phrasing
```

***

## Never Let the Model Own

```text theme={"dark"}
truth state
fake external execution
unverified completion
contract shape
system identity
```

***

## Runtime Boundary

Model output must be refined into Lucia’s stable structure before reaching the operator.

GPT-5.6 Sol response refinement sits over verified context packs. It can explain why a recommendation matters, answer short follow-ups, and keep the operator calm. It cannot change the booking, route, payment, task, save, or completion truth supplied by the deterministic layer.

The current Engine makes the execution path explicit in `/admin/operator-focus` response metadata:

```text theme={"dark"}
configured_model = configured default or override
resolved_model = provider-returned model when directly observed
model_invoked = whether any model task was attempted
deterministic_path = whether the response used no model task
fallback_used = whether fallback behavior was used, independently of model invocation
```

These fields describe what happened for that response. The configured model alone must never be treated as proof of model invocation, and historical runs without captured provenance must not be retroactively labeled.

Current payment reasoning boundary:

```text theme={"dark"}
Stripe financial truth may suppress payment blocker attention when confirmed paid.
Property payment policy truth is not durable Engine-owned yet.
LIEA does not yet make full policy-aware due/overdue/partial/pending judgment.
Signal Stream is not yet wired to LIEA.
Admin renders payment truth read-only and does not mutate it.
```

If model output fails, Lucia should degrade safely.

[06 - Semantic Conversational Intent Assist](/lucia/intelligence/semantic-conversational-intent-assist) supports this boundary by helping Lucia classify lightweight utility and conversational prompt families before response shaping. It does not own truth-state, external execution, or open-domain general assistance.

Core ordering:

```text theme={"dark"}
Human first. Purpose second. Boundary third.
```

***

## Workspace Context Reasoning

Lucia can now reason from the operator's current workspace surface.

Current Development behavior:

```text theme={"dark"}
Engine normalizes active_context.workspace safely
current surface context can orient the answer
explicit prompt subject still wins over page context
short next-step prompts resolve through workspace and prior context
saved DAW workflow state can inform the next response
```

Supported surfaces include:

```text theme={"dark"}
Dashboard
Calendar
DAW
Full Booking
Bookings
Maintenance
Concierge
Tasks
Reconciliation
```

Reasoning must preserve truth-state. A saved DAW step can be acknowledged as saved, but it must not become a claim that the payment, arrival detail, guest request, or underlying issue is resolved.

Operator-facing reasoning must not expose raw implementation language such as active\_context, current\_surface, payload, or metadata.

***

## Good Reasoning Behavior

Lucia should say:

```text theme={"dark"}
This matters because check-in is tomorrow and the issue was guest-reported.
```

Not:

```text theme={"dark"}
This appears to be an operational concern that may require review.
```

***

## Model Upgrade Discipline

Model changes are allowed and expected.

But the following must remain stable:

```text theme={"dark"}
response contract
truth-state behavior
operator calm
routing correctness
permission-based action language
```

***

## See Also

* [09 - OpenAI Model Layer](/lucia/infrastructure/openai-model-layer)
* [02 - Truth State Doctrine](/lucia/foundation/truth-state-doctrine)
* [04 - Lucia Workspace OS Milestone](/lucia/system-architecture/lucia-workspace-os-milestone)
* [00 - START HERE - Evaluation System](/lucia/evaluation/start-here-evaluation-system)