Skip to main content
This is the current-state spine for HelloLucia as of 2026-09-10. Unless a row or claim explicitly names a different date, the page-level as_of and last_verified dates apply to it; an explicit date overrides only that row or claim. A repository default, configured model, package version, successful test, or reachable URL proves only that fact; none of them silently proves a different surface, deployment, or future release. A Linear status is never evidence: every dated claim below names a commit, a served identity, or a recorded founder ruling. Work that is still at its gate or awaiting the founder’s verdict is written as such, not as settled truth.

​
How to read this page

Current truth has three independent axes:
See Truth State Doctrine for the durable vocabulary and Canon Update Ritual for the update rules, including the truth_scope controlled list.

​
Fleet at a glance

The three Mintlify products—Lucia, Lucia Fieldwork, and Eval Labs—describe this documentation site’s navigation. They are not a complete count of repositories, deployed services, or company operating surfaces.

​
September arc (2026-09-02 to 2026-09-10)

The rows below are dated by the record that carries the founder’s ruling and by the commit or served identity the record names. Where a ruling is built on the Development services but the founder’s verdict is still open on Linear, the row says so; nothing here is promoted to settled truth by a status.

​
Exact evidence ledger

These identifiers are evidence anchors, not a fleet-wide release number. A newer accepted deployment receipt supersedes its row without changing durable doctrine.

​
Current release identity

HelloLucia no longer has one meaningful company-wide runtime version. The Engine package still reports 0.1.3.6 on 2026-09-10, but that value spans many commits and deployments. It must not be used as the identity of Admin, Fieldwork, Guest Agent, Eval Labs, Workbench, or the company as a whole.

​
Model ownership

The durable rule is unchanged: deterministic systems own facts, identity, authorization, contracts, side effects, and verified completion. Models may assist interpretation, judgment, explanation, and tone inside those boundaries. See OpenAI Model Layer for the gateway page (being renamed Model Layer in Canon Wave 3).

​
Product and ownership topology

The canonical repository is the source owner. Filesystem location does not establish authority. Inside the Engine, the Algorithm files are owned by one engineer and every behavior change carries a plain-English before/after (see Lucia thread starter and task routing).

​
Environment and routing posture

​
Admin and Engine

  • Development services track the active dev branches and deploy on push once their repository gates pass; each runs one instance and no migration step runs on deploy. Engine migrations are hand-applied; the automatic step is planned, not built (Engine migrations ride the deploy, LUCI-146).
  • Staging services exist on the separately promoted stage branches. They are not aliases for the current Development commits.
  • Every /admin/* read and write requires the verified Clerk operator session; an anonymous request answers 401. Built and verified on Development (Admin operator routes take unauthenticated writes, LUCI-171, 2026-09-03); founder verdict open.
  • The generic maintenance webhook requires the shared X-Fieldwork-Secret header (Webhook accepts unauthenticated writes, LUCI-169, completed 2026-08-30).
  • Engine Development implements the Twilio and Postmark intake routes; Postmark is the employee-facing maintenance thread only.
  • Outbound guest mail leaves through Resend: the booking magic link and, since 2026-09-09, the payment request; SMS stays on Twilio. On Development every guest mail is rerouted to one testing inbox by a sink setting; a production service with the sink set refuses to boot.
  • Eval Labs evaluates the configured Engine target and must persist endpoint/environment provenance for new evidence.
The old sentence “the company baseline has not been promoted to Staging” is retired. Promotion is a per-repository, per-commit, per-environment fact.

​
Fieldwork

  • luciafieldwork.ai is the employee App boundary, served by Cloudflare Pages from main.
  • api.luciafieldwork.ai is the authenticated Fieldwork API and Postgres workflow boundary.
  • api-dev.hellolucia.ai is the Engine Development service dependency; the browser does not call Engine or a model provider directly.
  • Clerk organization-to-property authority controls application access. Older one-time-code and Cloudflare-Access-as-app-auth descriptions are historical.
  • A submitted item’s message thread is durable API-owned state, separate from the draft conversation. The App presents it; Engine relays it into the Maintenance Inbox; Admin renders operator context. No participant may fabricate delivery, seen, or sender identity.
  • The draft routing law (founder, 2026-08-29): a draft never reopens into chat; an edit always opens the manual form, and any exception needs the founder’s explicit yes (Leaving chat saves and clears, FILD-133).

​
Guest Agent

  • guest.hellolucia.ai is a Netlify production deployment context for a controlled-preview product.
  • Hosting context and the word “production” do not by themselves establish broad product readiness.
  • Booking verification and optional operational signals cross into Engine through explicit server contracts; Guest generation remains Guest-owned.

​
Eval Labs and separately owned tooling

  • evaluationlabs.ai is the production internal evaluation/evidence surface.
  • Workbench is separately owned private operator tooling outside this public Library.
  • Fleet Control Scanner is candidate source and is not deployed.
  • A Symphony-derived executor remains an architectural candidate and is not installed or deployed.
See Domains and Live Routing for the detailed domain matrix.

​
Evaluation truth

Eval Labs remains first-class Lucia intelligence infrastructure. It owns run orchestration, persisted evidence, review workflows, role scope, analysis surfaces, and Runtime Registry observations. Human reviewers own behavioral judgment. Engine and Guest Agent own their runtime responses and direct model proof. Historical scores such as 47/47, 178/178, and the 60-run / 3,000-prompt platform-readiness gate are dated evidence about exact suites and revisions. They are not today’s fleet-wide baseline and they do not prove human approval. The Brain quality eval bank (217 prompts) is not green and never was. The 2026-08-12 reading of 76/217 was a non-strict diagnostic; the Algorithm Scientist’s 2026-09-03 runs on dev recorded 174/217 and 192/217, and the standing failures are triaged on their own record (Eval bank has standing failures, ALGO-36). Until that bank is green, the honest gate for a language or copy change is identical failing sets before and after, zero new failures. No green baseline is claimed here. The standing behavior gate for the Algorithm is the replay harness: the fixture corpus runs through the old and the new Algorithm and the diff posts to the merge record as a plain table, with “no behavioral change” as the explicit happy path (Algorithm changes show their effects, HLLC-29, completed 2026-09-03). The contract conformance suite (One honest Algorithm contract, ALGO-8) and the chart-drift check ride the Engine’s npm test. Current evaluation areas include:
  • operator-facing Focus with Lucia and operational-attention behavior;
  • Guest Agent identity, verification, privacy, tone, and signal routing;
  • Fieldwork conversation/draft contracts, provenance, and behavior evaluation;
  • payment and other truth-state regressions where an owning runtime contract exists;
  • owner-language and names-law pins in each repository’s battery;
  • platform lifecycle, persistence, access, and review-evidence integrity.
See Evaluation System, Eval Labs Platform, and Validation Battery.

​
Villa Valentin and fixtures

Villa Valentin remains Lucia’s founding client and real-world proving ground, and it is the only property: “Everything is Villa Valentin right now. There is NO OTHER PROPERTY.” (founder, 2026-08-04; Pilot scope binds intake, LUCI-92). The canonical slug is villa_valentin, and every maintenance record binds to it before it enters the inbox. Since 2026-09-02 the Development fixture is no longer a static seed. A private synthetic world, lucia-world, plays the Villa on the real property clock against the Development services through the platform’s real doors, in Stripe test mode only, with its day zero re-anchored by reseat (“This needs to be in constant movement.”, founder, 2026-09-02; World fixture becomes time-relative, ALGO-23). It is an instrument for the Algorithm and the evaluators, never a product surface, and its internals are not published here.
All three must remain distinguishable.

​
Documentation state

The canonical public Library is:
The former Obsidian Publish hosts are retired. docs.hellolucia.ai and docs.evaluationlabs.ai have answered with a permanent (301) redirect to https://helloluciallc.com since 2026-08-15 (Legacy docs redirect home, HLLC-24); both roots were re-read on 2026-09-10. Workbench documentation is separately owned and outside this public Library.

​
What is not claimed

  • No DNS, domain, provider, credential, database, runtime configuration, or product service changed in this documentation update.
  • A configured model is not proof that a model ran, and a provider pin on one task is not a fleet-wide default.
  • A production hosting context is not broad product-readiness proof.
  • A passed historical suite is not permanent behavior proof, and the Brain eval bank is not green.
  • A saved workflow step is not the same as a resolved operational issue; the truth closes the step.
  • Work at its gate or awaiting the founder’s verdict is dated and named above; it is not settled current truth, and Stripe live mode is not on.
  • A roadmap, blueprint, or scanner candidate is not an installed autonomous executor.

​
See also