Solomon only earns the word “verification” if we apply it to ourselves, so we work in gates: a named set of promises, a date, and a pass where each one either holds or doesn't. This page is the current gate, in plain language, with an honest status on every line — including the lines that are not met.
It is the forward half of the same two questions Solomon has to answer for a technician: how do you know, and what did you check. The measurements page answers the first for work already done. This page answers the second in advance — here is what we said we would check, and how far along it is. Publishing it before the date is the point; a roadmap you only see after everything worked is not accountability.
Turn the ruler into the product's spine. The rules an answer must follow become one written file that both the overnight measurement and the live product read, and every result becomes a dated record anyone can download and re-check.
Opened October 7, 2026 · verification target: the week of November 17, 2026 (proposed)
The rules an answer must follow live in one versioned file. The overnight measurement scores against it, and the product enforces it while answering. Proof: the same rule shows up in a nightly result and in a live answer.
Each run, verdict, comparison and correction is stored as a dated record. The measurements page is generated from them rather than typed, and a correction is a new dated record, not an edit.
The records export in a standard, documented format. Each one carries a fingerprint and says which site, machine and run it came from.
No relevant documents means Solomon says so instead of answering. A source it could not read is reported as a failed read, not as an absence. Every part number, section or value in an answer must appear in its sources. Every machine's time budget is enforced. Each check ships with a test showing it catches what it is for.
“Not in this device's documents”, “couldn't check”, and “two sources disagree, here are both” each appear in the product as their own answer, never folded into a generic one.
Conflicts between sources, reworded, reordered and repeated, at least 200 scored runs, published with the count, the number of repeats and the limits of what it covers.
Every page and document changed during the gate is checked against our list of withdrawn figures before it ships.
Updated at gate events and weekly reviews — not on every change. If a status here looks stale against the measurements page, trust the measurements page.
Corrected October 7, 2026 — Until today this page still showed Gate C as the current gate, more than two weeks after it closed on September 21, and still listed the veteran's-notes result as met after that result was withdrawn on October 5. It now shows Gate E, and Gate C's entry below carries its correction.
The architecture proved out: knowledge captured in conversation, synced to the device, and answered offline with cited sources — verified live across six scripted beats.
The product became self-evident: structured answer cards in chat, a one-command install, and an evaluation harness that gates every model change.
Demonstrable anywhere, measured continuously: the engine ran offline on a single-board computer that a visitor could use from their own phone, the optional cloud half worked end to end, a cold machine reached a live demo in under five minutes, and the measurement ran itself overnight, unattended.
Corrected October 7, 2026: this gate also counted “when the veteran's notes contradict the manual, the notes win” as met. On October 5, 2026 that result was withdrawn: the scorer behind it was found to credit answers that led with the manual's fix. It is being re-measured, and it is not claimed until it is.
A second, unrelated open model ran through the same unmodified measurement; the device began recording which of its documents amend or supersede which; and the founder's own material went through the same pipeline as a technician's notes. Seven items met, one met as a draft, two carried into the next gate with their evidence, none abandoned.
A roadmap is easy to write and easy to quietly revise. Ours sits next to the numbers, carries a date, and says which promises are only partly kept — so the next version can be checked against this one. A gate closes here only when the evidence for it exists on the measurements page, not when we decide it feels done. If a line stops moving, that is a fair question to put to us.
Ask us about a line →