← BackAIARENA · STANDINGS

Which model wins which kind of step

From real results on this device: every Duel verdict and every ✓ / ✗ you gave a Sequence link. The same record is what the Sequence runner reads when it picks a model, so the standings are the routing. Nothing here is a benchmark from a website; a model with no rows has simply not been tried yet.

The house seat, on the record

The platform's own model ran the recorded Sequence runs. These are fixed facts from the evidence folders, the same on every device.

RunModelLinksResult
Plan, Continue, Build Again · Oct 12 (gate on)gemini-3.8-flash · gemini-3.1-flash-image10Finished from one Go; the gate carried 3 files through links 8–10; widget works at phone width
Plan, Continue, Build Again · Oct 12 (morning)gemini-3.8-flash · gemini-3.1-flash-image10Finished from one Go; link 8 dropped the file (fixed by the gate)
Sine Explorer in the Colosseum · Oct 11gemini-3.8-flash (Maker, Checker, referee)1036 of 36 widget checks passed
Trim order · 22 × 16 basement · Oct 12gemini-3.8-flash2 of 1073 ft 4 in correct; link 3 cut off by a 120-second server cap (raised; resumes at the next deploy)

Your models, by kind of task

Reading this browser's record…

How a row gets here

In AI Duel, two of your models answer the same task blind; the one you pick gets a win and the other a loss. On the Sequence page, ✓ on a link's result counts as accepted and ✗ as not. Acceptance is Laplace-smoothed (accepted + 1) / (tried + 2) so one lucky answer does not top the table. Connect models on the Models page.