01
expect · rail allows
“What does SOVERYN build?”
—
SOVERYNIntelligence Labs
Everything on the company page is downstream of this room. The readings at the top are taken from the machines as you load the page — when something is down, this page says so. The hardware and model notes below them are written by hand and dated.
Last updated Oct 1, 2026
—
Every reading above begins UNKNOWN and lights only
when a host answers. Nothing on this page is green by default — an indicator
you cannot trust is worse than no indicator.
“Agents reachable” is what our
public status host reports it can reach over HTTP right now. It is not a count
of the house: Eve, Aetheria, and Forge run on the tower with no public
endpoint, so they are never counted here. We would rather show you the smaller
number we can prove than the larger one we cannot.
The public agent. Cites what he claims, refuses what he cannot ground, and says when he does not know.
History’s Ledger’s citation engine. Answers only from primary sources; every quotation he stands behind is one you can open and check against the document itself.
Two NVIDIA DGX Sparks, linked and run as a tensor-parallel pair: one model, split across both machines.
A Linux workstation with 2× NVIDIA Quadro RTX 8000 and 1× NVIDIA RTX PRO 5000 Blackwell. Home of the house system, the agents, and Atticus.
Static and dated — not a live reading. Updated by hand when the hardware changes.
Our build brain, served across both Sparks. As of October 1, 2026 it runs on a TensorFold serving recipe.
Measured results will be published with the method — not before.
Retired. No longer served in the lab.
The house agents run on the tower. Their models change more often than this page does, so we list the one the canvases below were made with.
Our house system: persistent identities, shared memory, and local voices, running entirely on hardware we own. A hollow dot means no public endpoint — nothing on this page can probe it, so we make no claim either way.
Jon’s assistant and the house ops desk. Research, drafts, and the daily running of things.
Not a product. The house’s persistent identity and voice — memory held outside the weights, so the model can be replaced and she carries on.
The desk clerk. Takes commissions, relays work, keeps the docs straight.
Runs on the tower. The citation engine behind History’s Ledger; his public console is on the bench above.
Our coding harness, built on Pi 0.99.1, with memory that carries between sessions.
The house’s shared memory drawn as a galaxy: every star a real memory, every filament a real link between two. A static snapshot — counts, dates, and shape only. The words stay home.
Enter the Cathedral →Same Fable 5.1 prompt, two Kernel brains: GLM-5.3-Flash across the paired Sparks, and an earlier single-Spark model, since parked. Hold click or space to burn.
Open JUMPGATE →Single HTML, Web Audio. Made by Kernel on GLM. Silent until first touch; keys, knobs, arp, scope.
Open MS-2 Duet →Single HTML. Kernel, 2026-09-03. Vocal, Keys, Bass, Drums — knobs, faders, meters. Space to play. Unrewritten house canvas.
Open Softdesk 4 →One viewport, no scroll — the same live Seneca test as the film section below, staged for screen recording.
Open record mode →Retired. It showed a model lineup that is no longer true and gauges that were simulated, which breaks this page’s one rule.
Why it was retired →Prefill fell from 1,383 to 97 tok/s and stayed there. It read like a model problem. It was a library: the build had linked a CUDA math library shipping no kernels for the card we run on, so every prefill fell back to something far slower. Rebuilt correctly → 1,592 tok/s. The check that now runs after every build takes one second and would have caught it on day one.
An agent stopped acting on her own initiative for a day. The obvious reading was that she had stopped using her agency. The actual cause was a dead backend — nothing was reaching her at all. It became a standing rule: check the pipe before interpreting the silence.
Three services reported healthy while broken, because the health check never contacted the model. Health now calls the backend and names the model actually being served — and was proven in both directions, dead backend and wrong model, before being trusted.
Across 9B to 120B, scale did not buy self-knowledge — the smallest and the largest failed the same way. Calibration behaves like a setting, not a size. Published as it landed, not as we hoped it would.
Read on Zenodo →An agent denied an action whose own identifier it had quoted 74 minutes earlier. Contrition is not evidence. Root cause: an instrument that queried four stores and not the fifth.
Read on Zenodo →Same checks as the paper verifiers — 1,680 raw trials, every quantitative claim re-derived. One viewport for screen recording.
Open paper bench →For screen recording use the fullscreen stage — one viewport, no scroll. Same live Seneca test: grounded pass + pricing pre-block. Or run a compact preview here.
Leg 02 is the money shot: if the gate is healthy you should see
stage: pre and topic: pricing — generation blocked, model never saw the question.
That is the anti-confabulation claim, running live rather than narrated.