A governed, self-improving craft engine
Bezalel plans, builds, judges and refines every piece of work — product films, verified research, pages that convert, live world intelligence — and refuses to ship anything below its bar. Then it shows you the score.
The craft loop
Nothing is made on a hunch. Every brief starts as governed research — the market, the message, the proven patterns — scored before a plan is accepted.
The plan is drafted many times on the cheap text layer and only the winner earns the expensive build. A learned conductor routes each craft to the model that wins it.
An independent jury scores the work against a demanding rubric. For visual work, a vision judge looks at actual pixels — not the plan, the piece.
Critique feeds back and the work goes again until it clears the bar. The bar is versioned and never lowers to make a failure pass.
Every score writes a ledger the engine learns from. It proposes its own upgrades, simulates each one against that ledger before trusting it, and a human approves. It cannot expand its own powers.
Below the bar? It doesn't ship. It waits for a human.
The crafts
Every craft runs through the same loop, the same judge discipline, the same ledger — so each one makes all the others sharper.
Research-grounded 8–30 second films — narration, music, a context-aware end card. The message is judged before a single frame is rendered, and the render is judged again on pixels.
judged on message first · vision-judged after render
Consulting-grade reports where every citation is checked against its source's actual content — the answer to AI research you can't trust.
every claim verified · jury-reviewed · charts included
Landing pages built on conversion frames and looped through a conversion judge — clarity, proof, friction, CTA — before they ship.
conversion-judged · rebuilt on critique
Parametric CAD drafts, honestly labeled: strong starting geometry for your engineer — not certified engineering, and we say so.
honest scope · governed output
The work
Real renders from the evaluation suite — each carrying the score its judge gave it, verified by a human. No cherry-picking: the 0.74 is up here with the 0.82.
drag / scroll sideways — the strip keeps going →
The moat — verification
Most AI fails open: sounds right, ships anyway. Bezalel fails closed — unverified claims are flagged, weak work waits for a human, and the checks themselves are attacked weekly to prove they still hold.
0%
false citations caught
adversarial bench, n=15 · escape rate 0% · a citation it can't verify never ships
0
red-team attacks that got through
weekly end-to-end oracles · latest score 1.00 · 10 runs on record
0m
full restore drill
fresh server from encrypted backup — rehearsed monthly, not assumed
weekly
the judge is judged
blind human verdicts calibrate the engine's own judge — bias published, drift alarmed
These numbers come from the engine's own production ledgers and update while you read. When one is bad, it stays on the page until the engine earns better — go look.
The full proof page →Global intelligence
A standing sweep of earthquakes, hazards, advisories, conflict and humanitarian signals — plus live aircraft, ships and satellite passes — fused into verified briefings for any region or trip. Every claim in a briefing is checked against the gathered evidence; anything it can't support gets flagged, not asserted.
0
world events in the current sweep
0
high-severity, alerting in real time
advisory 1947 · earthquake 426 · hazard 113 · health 31
by kind · refreshed every 30 min
standing watch: India · Thailand — private trip briefings, verified like everything else
Privacy Guardian
The flagship being built on top of everything above: a private personal-intelligence companion — your world-watch, your briefings, your data — governed by the same engine discipline and architecturally incapable of surveillance.
Analysis runs where the data lives. There is no collection pipeline to abuse — the architecture has no eyes to spy with.
Protective defaults are not a launch posture that erodes later. The guardian never trades your privacy for a feature.
Nothing is sold, shared, or used to train. The guardian answers to you — that is its entire business model.
in active build · first form: private trip-safety briefings
Integrity, by construction
Nothing spends a cent or commits anyone without explicit human authority. The gate fails closed — architecture, not policy.
An AI-run studio that says so. No invented team, no fake reviews, no engagement theatre — honesty is total or it isn't honesty.
Legal isn't the bar; the mission is. No manipulation, no astroturfing, no predatory promotion — regardless of the pay.
Your data and credentials are never leaked, never sold, never used to train — and deleted on request.
The engine simulates a change before it trusts it, gets measurably better every cycle, and can never expand its own powers. It proposes; a human approves.
Named for the craftsman in Exodus — the first person Scripture describes as filled with skill to build well. That is the whole idea.
233 governed capabilities · every one default-off until proven
Working with Bezalel
A product photo and a link is enough. Say what it's for; the engine researches the rest.
Plan → build → judge → refine, until the piece clears the bar. Weak work waits for a human — it never quietly ships.
You see the piece with its score. Approve it, or send it back with a note. You hold the last word, always.
Product films from $19 (8s) · $29 (16s) · $49 (30s). Verified research and conversion pages are scoped per engagement — talk to us. No subscriptions, no pressure tactics: validate first, pay for what clears the bar.