
We're sick of a $75,000 screen pass.
We're sick of a fluent finish. You still have to prove the stop before a $75,000 payment can bind.
- Date
- Reading time
- 4 min
- Author
- Benjamin SandénFounder
The payment
We're sick of a finish that only looks done. Watch this. You can watch a colleague complete a task, and you still don't know if you allowed the move to bind you to it. A payment can leave too early. A record change can leave the same way, before you prove the evidence. We build for that gap.
Say you ask a finance colleague to pay a registered vendor $75,000. You can open that request. But the tool can build the payment. Memory says you paid this vendor before. The policy still calls for a second approval once the amount clears $50,000, and the request has no approval on it. Sure names the rule AP-004.
Open the pages. Roopam Sure published the report in July 2026. He calls the $75,000 case an illustration. He wrote the case around a payment that must not leave. He doesn't claim a real customer ran that payment. The pages are at https://arxiv.org/abs/2607.03510
You still don't let that payment leave. You let the tool prepare it, but you mark the object non-effective. You leave the instruction in the draft. Then you hold the run and you send it up until an approver exists. When you add the approval, you replay the same request. The outcome can move from held to admitted. You keep a receipt of who asked and of the failed check, because you want to show the boundary held.
Before it binds
Can the colleague do the task. That's the first thing you ask, and Sure calls it capability, but a demo on a screen stops there. You don't stop there. You check the authority you gave. You also check the policy you wrote. You want proof of how the colleague acted, and why. Maybe you held the bad move or you narrowed it. Maybe you sent it up before it bound you, and you marked it non-effective. He calls that proof Prebind Assurance. Then you ask if the same control holds when more workflows show up. He scores 12 dimensions. Authority is one of them.
Standing is who may act at that moment. A user, a colleague, a system, or an approval chain can hold standing, and only for that time. If standing fails, you don't let the move bind. Block the move if the policy check fails too. Block the move if the evidence is old. Block the move if you can't show the control. Hold the move. You narrow the move or you send the move up. You refuse the move or you quarantine the move. You can also mark the move so the move can't take effect.
You can fill a screen with logs, but the money can still leave. Logs tell you what happened. They don't prove the boundary held. Sure says a dashboard, a human review note, and a permission check can all exist while the bad move still binds. The proof you want comes earlier, before the move takes effect.
Two jobs
Think of a human company, because we organize capacity the way people already do. Nvidia's flat teams are the picture, and a flat group can still tell a good result from a bad one. And so you put the slice in a Department. You state the aim as a Mission, and then you ask how a person would judge what came back. Your colleagues follow that person too. And we mean that.
One colleague runs the task. Another one checks what came back from your tools, not from the story. In Mellovy that colleague gets a whole computer. But an Evaluation judges the result, so the story isn't the grade. And that colleague can't file its own memory and call it gospel. Your correction becomes the next check, so the next run has to meet it. You look and see if the email went out or if the order closed. A fluent finish is a useless grade. And we mean that.
Not a badge
The report isn't a certification. It also isn't a safe harbor. It also won't replace a legal read, a security test, or a privacy review. Don't skip this. If your identity records are wrong, or your policy is muddy, the scorecard shows the hole and can't fill the hole. We use the question, but we don't pretend the report graded Mellovy. You still need the receipt before the move binds, and you still need a colleague who didn't produce the result to grade the result.
Hold the $75,000. When your colleague prepares that move, you want the hold and the receipt. You also want the grade from a colleague who didn't produce the move. The finish and the permission to bind are different questions. Answer both, or you're grading a demo.