Status
Where Varra and Lucentra actually stand today — updated as things genuinely change, not on a marketing schedule.
Last updated
Live means it is in the product and you can use it now. Staged means it is built and tested but nobody can reach it — either it sits behind a switch that is off, or it exists in the engine with no interface on it — so it changes nothing for anyone today; the row says which of the two. Planned means it is not built. Nothing on this page is described as more than it is; where a layer is partly one and partly another, it says which parts.
Varra, by architecture layer
The nine layers are described on the overview, and they are numbered the same way here. This is where each of them actually is.
01 — Generalist core
Live expanding
Deployed as a real, publicly usable web application — not a prototype or a closed demo — backed by a real database, real accounts, and a genuinely tested codebase. Real-time streaming and prompt caching are both wired in. Image and PDF input work. An optional working memory lets the model keep its own notes across a session. Effort is selectable between faster and deeper.
02 — Router
Staged
Built, tested and carrying its own eval set and telemetry: deterministic rules first, a model classifier only where the rules are inconclusive, and every failure path — timeout, invalid output, telemetry error — failing open to the ordinary response path. It runs behind a three-position switch that reads as off unless deliberately turned on, with a shadow position that routes and records without applying anything. Nobody using Varra today is being routed. Model selection and risk scoring are design, not code.
03 — Work Items
Live, nothing automatic
The record that lets a piece of work be tracked rather than only answered: an account-owned item with a title, a description, one of four states — open, blocked, resolved, cancelled — an optional link to the conversation it came from, and versioned edits that reject a stale write. POST/GET/PATCH/DELETE /api/work-items exist in the engine, with database-backed tests, account export and fail-closed deletion. An outcome can be appended to one — achieved, partly achieved, not achieved, indeterminate. The app carries a Work Items view and a New Work Item control, so an item can be created, tracked and resolved by hand. What still does not happen is anything automatic: nothing creates a Work Item on its own, and no routing, verification or execution path acts on one. Saving is explicit, and resolved never means verified, approved or executed. The layer →
04 — Specialist Skills
16 modes live 6 Skill packs staged
Sixteen work modes ship in the product and you select them yourself. Separately, six Skill packs exist in the routing registry — construction invoice audit, document evidence analysis, electrical troubleshooting, HVAC diagnostics, financial analysis, investment research — declaring which others they compose with. They are behind the same switch as the router, so no one is receiving one. Full catalogue →
05 — Company context
Planned per-account reference data live
There is no company workspace in Varra. You cannot upload policies, connect a system, or set an approval rule. One narrow piece is live and in production: per-account reference data. GET/PUT/DELETE /api/reference back a “Your reference data” panel in the account menu, entries persist per account and per mode, and the precedence is account value, then built-in table, then the model's own memory. It applies on all four response paths, and an answer that used it is stamped as coming from the account. Labels are stored under the same canonical reduction lookups use, so an entry typed “12 AWG copper” matches the questions “12 awg” matches. The intended layer →
06 — Tools
Partly live
Live in the deployed product: web search, used when a question depends on current information rather than on every query. Also live: document and image input, and the session working memory. Built into the engine but not switched on: web fetch, code execution, and a stronger advisor model for hard questions. Not built at all: any connection to an external business system — email, storage, project management, accounting, databases. Varra connects to none of them.
07 — Varra Assurance
Limited live coverage
One mechanism works end to end: where an answer's key figure matches a verified table, the table's entry replaces the model's text and the answer is stamped with its source. Three of the sixteen modes carry such a table; two are published in full, at nine rows total. Most Varra answers today are therefore not independently verified, and Varra says so rather than implying otherwise. Coverage and limits →
08 — Varra Runtime (execution)
Planned early architecture
Varra takes no action in any external system. It holds no credentials to one, writes nothing anywhere, and submits nothing on anyone's behalf. There is no runtime service and no execution gateway in the codebase. What exists today is answering and preparing text. Broken down, so that no part of it is read as further along than it is:
09 — Outcomes and learning
Recording live, learning planned
No outcome feedback loop exists. Varra does not learn from corrections across users or sessions, and no model is trained or fine-tuned on customer conversations. The one piece of groundwork: an outcome assessment can be appended to a Work Item and read back as history, and the app now carries a form for writing one. Nothing consumes it — no routing decision, no model, no behaviour anywhere. Recording an outcome is not learning from one.
Testing
Including real database-backed integration tests rather than mocks alone, plus a separate eval set for the routing layer. The count is re-derived from the repositories rather than typed by hand, and it is stated as a floor because an exact number is wrong the next time somebody writes a test.
Lucentra
No Lucentra hardware exists and none is shipping. It is a surface for the same Varra intelligence rather than a separate product, and it remains at an early, pre-hardware stage: reference-design sourcing is in progress, and every capability described for it — navigation, translation, object identification, captioning — is intended, not available. The software groundwork it will depend on is built and validated in software today, ahead of any hardware commitment: it runs the same sixteen work modes, carries over 75 automated tests of its own in continuous integration, and its heads-up display is scored for daylight legibility against worst-case backgrounds on every change — a question most projects cannot answer until hardware samples arrive.
Billing
Varra is free. An account is required so conversations persist. No card is asked for and no card can be entered. The Pro and Business tiers exist in the app and on the pricing page as descriptions of what those tiers will include, not as things anybody can buy.
Recently
-
Rows 03 and 09 were corrected against the deployed product, which had moved past them. Work Items and outcome assessments both have an interface in Varra now — a Work Items view, a New Work Item control and a form for recording an outcome — so the rows no longer say there is no way to reach either. What did not change is the part that mattered in both rows: nothing creates a Work Item on its own, and nothing consumes an outcome. The entry below, written when neither had an interface, is left as it was: it was accurate on its date, and a log that gets edited to match the present is not a log.
-
Work Items were added to the architecture as layer 03 — the account-owned record that lets a piece of work be tracked rather than only answered. The storage, the API and the outcome history exist in the engine; there is no interface for any of it, so the row above says so plainly. The site's wording was tightened at the same time to describe Varra as a system that carries work to a result rather than one that returns answers, and this page's layer numbering was corrected to match the overview, which it had silently disagreed with. No capability changed.
-
Varra Runtime — the planned execution layer, covering connectors, browser, sandbox and desktop environments — was added to the architecture on the overview, business and roadmap pages, and broken out here part by part. Nothing about it is built: no runtime service, no execution gateway, no connector to any external system. The two pieces of real groundwork are named above rather than rounded up.
-
The site was rebuilt around what Varra is becoming: a general-purpose system with a generalist core, a router, specialist Skills, tools, company context and verification, rather than a product defined by its trade modes. Skills and Business pages were added, the reference page became Proof, and this page gained a layer-by-layer status breakdown. No capability changed — only how accurately the site describes them.
-
Every page got a skip link, so a keyboard user no longer tabs through the whole top bar to reach the content — and the body typeface was declared as the variable font it is, which restores bold emphasis that had been rendering at medium weight.
-
The homepage was rebuilt around what can actually be checked: the reference substitution is set as text rather than a screenshot, and a composer in the hero hands a typed question straight to Varra.
-
Varra moved to its own domain, and the skill-routing layer gained its first finance and document skills — still behind the same switch, still off by default.
-
Rate limiting and the spend cap moved into the database, so neither resets when the service restarts.
-
A skill-routing layer landed in the engine with its own eval set. It sits behind a switch that reads as off unless it is deliberately turned on, so it changes nothing for anyone using Varra today.
-
The first two verified reference tables were published on this site — generated from the engine's own source rather than transcribed, so the page cannot drift from the product.
-
The logomark shipped across every page, and link previews got a declared banner instead of a bare URL.
Company
Syntur was established in 2026. Syntur, Varra and Lucentra are working names pending formal trademark clearance.