Workspace · one real model failure
What makes this bird classifier fail?
Real benchmark data, computed in your browser.
How this page works. The partner is a scripted guide: its lines were written in advance. A plan stays in pencil while you can still edit it; inking it fixes the plan and starts its runs. Holds are limits you set on a plan: how far average accuracy may fall, how long the refits may run, and whether the test images stay sealed, kept from every run until you choose one option to score on them, once. Holds are enforced from the moment the plan is inked. The gate is a fixed rule: it puts a 90% interval on each option’s change in average accuracy and judges the option within your limit, beyond it or too close to call.
Loading the case and fitting the readout in your browser…