The sign-off can’t come from the factory that wrote the code. Right now it’s a person who never gets to go home. jusFactory makes it something you can issue and check.
A droid writes the code. The factory generates it. Validation approves what is safe and rejects the rest. Only approved work becomes value.
Today we certify the people who run the agents: a real exam, a public credential. The agent and the code come next, built on the same proof. The credential lives outside the factory, so agent-built software ships with a name behind it.
Rung one is live, in pilot with a handful of Upekkha portfolio companies. Engineers can take it for $25. The rest of the ladder is in build, in public.
Right now the bar for what’s safe to ship lives in your best architect’s head. When they’re out, the bar is out with them. New engineers learn it by guessing.
The exam records how a skilled person decides what to trust: what they accept from an agent, what they reject, where they stop and check. Every attempt turns one person’s judgment into something the rest of the team can be measured against. The standard stops being tribal knowledge and starts being written down.
The credential is what the engineer keeps. The recorded judgment is what the company keeps. It is the thing that lets the next rung exist, and the reason the bar no longer leaves when a person does.
Agents can write code. They cannot be trusted, coordinated, or audited, because the intent behind the code was never captured in a system.
Underneath all three: context lives in the wrong place. In a senior engineer’s head, a stale doc, a 2014 comment that says do not touch this. AI does not fix tribal knowledge. It amplifies it.
Same problem, a different door for each person who signs →A ladder, not a suite. Each rung is built from the one below it. You can’t certify the agent until you know what good human judgment looks like, and you can’t certify the code until the agent clears the bar a human set. So we built from the bottom.
A timed exam. 50 questions, 90 minutes, pass at 40 out of 50. It tests whether a developer can direct, correct, and check agent work. Pass and you get a public credential anyone can verify. $25 an attempt. This rung runs today.
Next. The judgments certified engineers make on rung one become the bar an agent has to clear. Once an agent clears it, you stop reading every result and spot-check a worker you already trust. We are still designing how an agent earns this. Not built yet.
The end of the ladder. Trusted code ships by default instead of by exception, with the check in the system and a name on every change. This is the hardest rung and the furthest out. We will not claim it works until it does.
We did not ship all three and call it a platform. We shipped the rung that works. The rest is on the roadmap, in public, where you can watch it.
If a verification tool oversells, you should not trust it. So we do not.
When something machine-written breaks in prod and someone asks who approved it, the answer has to exist. That answer is the thing we are building. The labs cannot make it, because the proof comes from human judgment they do not have. We are already collecting it, $25 at a time.
One exam, one attempt, twenty-five dollars. 50 questions, 90 minutes, pass at 40. Keep the credential. No seats, no annual contract, no demo gate. Buy one. Take it. See if it is any good.
You don’t yet know if your engineers are agentic, or if your factory is safe to ship from. We do, with proof. Start with the rung that’s live.