Two questions you cannot ask most AI. Solomon is the trust and verification layer that makes both answerable, and it does three things to get there: your knowledge stays on hardware you own; questions are answered on that same hardware, with no line back to anyone's servers; and every answer arrives with its sources attached — saying plainly when it could not check one, because could-not-check and not-there are different answers.
Its first use case is industrial operations. Verification there is already legally mandatory, the buildings are places the cloud cannot legally or physically reach, and the reader is a technician at a stopped machine deciding whether to trust an answer before acting on it. If the layer holds anywhere, it has to hold there first.
An illustration: surfacing the veteran's logged fix alongside the manual's procedure, with both cited.
Custody, on-site answering and checkable sources are general properties. Here is what they look like pointed at one problem — a retiring workforce, machines that stop, and a building with no usable connection. Three steps, each one a piece of the layer doing a job somebody on that floor actually has.
Veterans record what they know by typing it the way they'd say it. Manuals get dropped in as PDFs — layout, tables, and section structure preserved. Solomon confirms what it understood, names the machine and the fault, and indexes everything locally. Re-upload a document and it says "updated," not duplicated.
Squaloo's OB2B protocol moves knowledge between headquarters and the device beside the machine — differential, verified, and honest: every sync posts a receipt with real counts, and a failed sync says so. No cloud dependence, ever.
A technician asks in plain English. Seconds later: numbered steps, lockout first, citing the manual section AND the veteran's logged fix, side by side. No sources found? Solomon says so instead of guessing.
to a cited answer, measured warm by the eval harness calling the engine in-process — median 7.5s across 14 runs, 2026-09-05, machine not recorded. Not the ask page or chat, which add the request path and queueing
offline — verified with the network disconnected, 2026-09-05
the single-board computer's budget. Each machine now has its own — a laptop and a credit-card-sized computer are not the same promise (2026-09-27)
cloud services in the answer path — nothing to reprice, nothing to leak
These are early results, and we label them that way on purpose. They test one thing: whether the layer is viable for this first use case. Measured live, August–September 2026 — knowledge captured in conversation, synced to the device, the network disconnected, and a junior technician's question answered with its sources attached. Measured, not estimated, and each figure carries the date it was true.
What they do not yet show: a wide corpus, many machines, or many technicians. The scripted question set is small and we say so wherever we quote it. As the work gets more concrete these numbers change — and when they do, the old ones stay on the measurements page with their dates rather than disappearing.
We publish the full scorecard — including the questions Solomon still gets wrong, dated and linked to the fix. See the measurements →
Citations enforced by architecture — every answer carries its sources; traceability isn't left to the model's discretion.
Honest refusals — no documentation, no answer. Solomon never bluffs a repair procedure.
Compliance by design — data never leaves the facility, built for environments where that's non-negotiable.
Vendor-independent — open model weights we run ourselves. When one API vendor repriced overnight, we deleted the dependency in an afternoon.
How do you know? — every answer carries its sources, and says plainly when it could not check one, because “could not check” is not the same as “not there”. Where does it live? — your building, your box, your keys; cut the cord and nothing stops working. Who is checking? — not us. A published ruler, measured nightly, that anyone can re-run, including on us.
We believe the binding underneath works for any model — and we have scheduled the experiment that could prove us wrong. We will publish that result whichever way it goes.
Cloud connected-worker platforms promise AI teammates — with an asterisk: requires connectivity, expands your compliance boundary into their cloud, and rides on a model vendor who can reprice overnight. Solomon is everything they promise, minus the asterisk.
If any of this is useful to you — a facility with machines that cannot afford to forget, a question about how the verification actually works, or a claim on this page you want to check — write to us. We answer technical questions with the artifact attached, and we will run the demo live, on a machine with its network disconnected, for anyone who asks.
[email protected]