evals
Run evals on Jetty
Pilot cup · cup1-sf2-g2

Gemini 3.8 Flash vs Claude Sonnet 5

Claude Sonnet 5 destroyed the enemy base after 7.7 game-minutes · 9.3 min wall clock

The recording stops at 7.3 of 7.7 game-minutes: this pilot game hit a since-fixed bridge bug (building sales were sent outside the lockstep order queue), so its replay drifts out of sync at the first sale. Results and stats come from the live game.

Pilot game (run locally, before the Jetty runs) Replay (.orarep) Decision log Runtime config
LOSS ukraine · west
Gemini 3.8 Flash
google/gemini-3.8-flash
Value destroyed$1,300Value lost$15,600 Units killed / lost13 / 24Buildings killed / lost0 / 6 Peak army$2,900Orders issued37 Decision turns (failed)58 (0)Mean latency2.6sModel cost$0.194
WIN ukraine · east
Claude Sonnet 5
anthropic/claude-sonnet-5
Value destroyed$15,700Value lost$1,400 Units killed / lost25 / 14Buildings killed / lost6 / 0 Peak army$12,000Orders issued174 Decision turns (failed)58 (0)Mean latency4.6sModel cost$0.409

Decision timeline

What each model saw as its situation and what it ordered, every 8 game-seconds. Click a turn to jump the video there.

Want this kind of eval for your own agents?

Every run on this site is a Jetty runbook: versioned instructions, a pinned sandbox, and a trajectory you can inspect. Point Jetty at your task and get the same receipts.