Skip to main content
Lucia’s model layer is one Engine-owned gateway with two providers behind it. The provider is resolved from the model id a task is pointed at: a claude-* model goes to Anthropic, anything else goes to OpenAI. Fable is Lucia’s voice where a task is pinned to it; GPT-5.6 Sol remains the configured default.
This page was OpenAI Model Layer until 2026-09-10. The old address redirects here. The single-provider description it carried is dated history; the gateway has spoken both providers on Engine dev since 2026-09-04 (Fable is Lucia’s voice, ALGO-39).

​
Current posture

Which model a task uses is a setting, not a code path. The Engine reads one default (LUCIA_INTELLIGENCE_MODEL) and a per-task pin (LUCIA_MODEL_<TASK>) for each task in the source map:
The source comment states the rule plainly: “Anthropic model ids such as claude-fable-5-1 and claude-opus-5 are valid values.” Provider follows the model id; every caller of the gateway gets both providers for free.

​
What runs on Fable today

Two tasks are pinned to claude-fable-5-1 on the Development Engine, both with no fallback: The founder’s words that set this, on Fable is Lucia’s voice (ALGO-39): “I think Anthropic’s models should be her predominant voice.” (2026-09-03) and, after the first live Fable answers, “Keep Fable on.” (2026-09-04, 06:46 PT). On the same day the founder kept the Fieldwork brain where it was: “GPT is fine here and we can save token spend by keeping this on GPT and reserving Fable for what really matters, in this case, the core Lucia.” The no-fallback rule is the quiet-and-honest rule: when Fable cannot answer, Lucia says so in one owner-language sentence rather than letting another mind speak in her voice. Focus answers from one brain (ALGO-40) records the founder’s ruling as “No fallback: quiet and honest.” (2026-09-04). The source carries it as a per-task default that the literal env value none also expresses (LUCIA_FALLBACK_MODEL_<TASK>=none). For the Focus composition the env line is recorded set on api-dev (LUCIA_FALLBACK_MODEL_FOCUS_OPS_HOUSE_COMPOSITION=none, ALGO-40 gate, 2026-09-04). For the payment draft the equivalent line is not evidenced on any record and is treated as set by the founder’s hand, unverified; the source default already yields no fallback for both tasks at this commit. Everything not pinned runs on the default. The Fieldwork conversation brain and the guest-facing tasks stay on GPT-5.6 Sol; the Fieldwork conversation provider path is a separate, LIEA-contracted file that the gateway change did not touch.

​
The Anthropic path, as built

Verified in the Engine source at the deployed commit:
  • Messages API with the same JSON schema the task declares; the request asks for exactly one JSON object and nothing else.
  • No temperature reaches Anthropic: the request builders accept and ignore it after Anthropic rejected it as deprecated (ALGO-39 addendum, 2026-09-04); every Anthropic request is model, system, messages and max_tokens, plus the JSON output format when the schema passes the grammar audit. GPT-5 family requests omit temperature too (shouldOmitTemperature).
  • 45 s timeout for Anthropic calls by default (LUCIA_ANTHROPIC_DEFAULT_TIMEOUT_MS = 45_000, overridable by name through LUCIA_ANTHROPIC_TIMEOUT_MS). The source comment records why: Fable answers a house question in 5 to 12 s live and one composition died at the old 15 s cut. OpenAI keeps its own defaults.
  • Usage is read from the provider’s own response, never estimated. For Anthropic the billed input is the sum of input_tokens, cache_read_input_tokens and cache_creation_input_tokens; for OpenAI the reported input_tokens already includes the cached part. Numbers only, no pricing.
  • Credential is LUCIA_ANTHROPIC_API_KEY first, then ANTHROPIC_API_KEY, by name. Readiness reports the provider and whether its key is configured.

​
Failover, recorded

The gateway retries once on the configured fallback model when the primary provider fails with a transient class (timeout, 5xx, rate limit, overloaded, per the provider error classifier). It does not fail over on authentication errors, schema rejections, or a model refusal; those are reported as the task’s own error. Malformed output is not a failover trigger. A task whose fallback resolves to none never retries on another model. The Fieldwork conversation task never fails over. The global fallback is LUCIA_FALLBACK_MODEL (default gpt-5.6-sol); a task may override it with LUCIA_FALLBACK_MODEL_<TASK>.

​
Provenance: the ledger records the provider

Every gateway diagnostics record carries:
This is the founder’s “ledger records provider” law, carried into the gateway by Fable is Lucia’s voice (ALGO-39). The served Focus route exposes its path in meta.focus_path (house_pack or house_pack_unavailable); see Focus Ops Intelligence. The runtime-verification route reports which provider each task would use (getLuciaModelConfigSnapshot().providers). A 200 still does not by itself prove model execution or provider success; read provenance rather than infer it from the route status or the configured default.

​
What the model does not own

Unchanged doctrine. The model must not own:
Model output sits above deterministic truth. Guest identity, booking IDs, stay windows, payment state, task status, route and action safety, and completion truth remain system-owned. Everything feeds the Algorithm; the Algorithm decides what surfaces. See LIEA doctrine.

​
Not current

  • The predominant voice (ALGO-39 phase 03) was executed as a decision, not a flip (2026-09-04): the default stays gpt-5.6-sol, Fable is pinned per owner-facing task, GPT stays the Fieldwork brain and the recorded global fallback. No task-by-task flip of the default has happened.
  • A stand-in voice when Fable is down (ALGO-41) is Planned and is a recorded launch blocker; nothing stands in today.
  • ALGO-32 and ALGO-33 are Planned; the gateway carries the provider ledger fields, not those issues’ further scope.
  • Lucia Guest Agent is an independently owned runtime. Its last accepted same-deploy proof (2026-07-18, source fbf77b662511cf004ee1f46787e05773a8a78d71, configured and provider-resolved gpt-5.6-sol) is dated evidence about that surface only.
  • Staging still serves Engine stage commit 9a9227e8fb40e94748c160265a697994023361ca (2026-04-28). None of the September gateway work is promoted there.

​
Historical: the single-provider gateway

Between 2026-05-04 and 2026-09-04 the JSON gateway spoke only the OpenAI Responses API, with gpt-5.6-sol as the source default (recorded at Engine dev 9a15b1da199923d1ecb0927be2507feda2cb4c71, 2026-08-12). The legacy Focus path that used focus_ops_intent_assist and focus_ops_response_refinement left the served route on 2026-09-04 (ALGO-40); its task keys remain in the model map only where a caller still exists.

​
See also