Turn automation up only as far as you can prove.
A model can propose labels all day. The question is which of those proposals a person is allowed to skip. Nitsor puts that on one dial, and the dial stops at the point your own measurement stops. Past that point there is nothing to back the decision up, so the dial will not go there.
Every proposal waits for a person until a measurement says it need not. We take that measurement with you, on your own scans.
Turn the dial. Watch it stop.
Sixteen proposals on one published CT slice. Drag the dial to say how many of them may skip a person. It moves freely up to the measured bar and refuses to go further, because past the bar nothing has been measured.
each square is one proposal on this slice · orange means a person looks at it · bone means the measurement covered it
real scanThe sixteen proposals are the instances the dataset publishes on slice 192 of the Fraunhofer EZRT XXL-CT Me 163 subvolume V5, doi:10.5281/zenodo.10651746, CC BY 4.0. The 62% bar is a stand-in. It represents a guarantee measured on one set of scans. On your data the number would be different, or there might be no number at all. The counts are arithmetic on those sixteen.
How the bar is set
The model proposes, nothing is saved
Predictions arrive as ghost proposals on the model's own branch. A person accepts or rejects each one, and the accept is what saves it.
Proposals are committed on a model branch with the model version in the record. They never enter the accepted set without an accept event carrying a human author.
Those decisions become the measurement
Your reviewers' accepts and rejects are the answer sheet. We use them to work out how often the model would have been wrong.
The measurement runs on your own scans, with your reviewers' decisions as the answer sheet. What counts as one error, what unit is counted and what the method assumes are all written down, so an auditor can disagree with them in writing.
The measurement produces a bar, not a score
A confidence number out of a model is not a measurement. The bar comes from checking the model against cases it never saw.
Distribution-free risk control, not softmax. The certificate names the risk bound it holds to, the model version and the data commit it was measured on, and the sample unit it holds over.
The dial cannot pass the bar
Turn it as far as the measurement covers. Past that, the control stops. There is no override that hides the missing evidence.
A proposal skips review only where a valid certificate covers its class and its conditions. Recalibration issues a new certificate; it never edits the old one.
Questions this has to answer
| Question | Answer |
|---|---|
| Why did nobody look at slice 412? | Because a certificate covered it. The answer resolves to a certificate number, a model version and a data commit, and you can read all three. |
| Can I turn it all the way up? | Only if the measurement goes that far. The dial stops at the bar, and the bar comes from your data, not from a setting. |
| Is the model's confidence the bar? | No. A confidence score is a number the model prints about itself. The bar comes from checking the model on cases it never saw. |
| What happens when the data changes? | The old certificate stops covering the new conditions. Automation falls back to review until a new measurement is made. |
| Who can move the dial? | Whoever you grant that, and the change is an event with a name on it, like every other change. |
What the dial is really for
- It stops where the evidence stops
- The dial will not turn past the point your own measurement reaches. That is the whole design. A tool that lets you turn it further is selling you a feeling.
- Nothing is saved by a model alone
- Proposals sit on the model's own branch with the model version on them. They join the accepted set only through an accept event carrying a person's name.
- The measurement is yours, not ours
- It runs on your scans and your reviewers' decisions. A number measured on somebody else's data would tell you nothing about your parts.
We do not publish a speed multiplier, because we have not measured one. If you report against a standard, tell us which one on the call and we will say plainly whether we help.