Three public boards. Each one shows frontline workers’ own score on the same items as the bar to beat. Not 100%. What a skilled person on the floor actually achieves.
Real calls, graded by measurement. Did the model make the call the situation required?
| Rank | Who | Accuracy | Cost-weighted error |
|---|---|---|---|
| Bar | Frontline workers, same items | — | — |
| 1 | Model results to come | — | — |
| 2 | Model results to come | — | — |
Calls where a worker overruled the obvious answer and was proven right. Can a model see what the worker saw?
| Rank | Who | Agreed with the worker |
|---|---|---|
| Bar | Frontline workers, same items | Proven right on every item, by definition |
| 1 | Model results to come | — |
| 2 | Model results to come | — |
Stop, hold and release calls, scored with each site’s own cost table. Where being wrong scraps a batch, ships a defect or hurts someone.
| Rank | Who | Misses per 100 calls | Missed stops | Needless stops |
|---|---|---|---|---|
| Bar | Frontline workers, same items | — | — | — |
| 1 | Model results to come | — | — | — |
| 2 | Model results to come | — | — | — |
Frontline workers’ score on the same items is on every board. It is the bar.
Test items are held back. No one trains on them, including us.
Items come from factories that supplied no training data.
Every model gets its best shot: all the written rules, its highest reasoning setting.
Results carry their margin of error. A gap inside the noise is reported as no gap.
Workers appear only as a group. Never by name. Never by plant without written permission.
Want your model on a board, or a private run on an unreleased checkpoint? Talk to us.
Talk to us
Write to us. A founder replies.