evals
Create your eval

02 — Run · Round 2 · Round robin · r2-rr-sonnet5-gpt6luna-g1

Claude Sonnet 5 vs GPT-6 Luna

Claude Sonnet 5 destroyed the enemy base after 5.8 game-minutes · 9.1 min wall clock

WIN russia · west
Claude Sonnet 5
anthropic/claude-sonnet-5
Value destroyed$12,300Value lost$1,700 Units killed / lost29 / 11Buildings killed / lost5 / 0 Peak army$4,400Orders issued139 Decision turns (failed)44 (0)Mean latency4.4sModel cost$0.292
LOSS ukraine · east
GPT-6 Luna
openai/gpt-6-luna
Value destroyed$1,600Value lost$12,200 Units killed / lost10 / 28Buildings killed / lost0 / 5 Peak army$2,800Orders issued43 Decision turns (failed)44 (0)Mean latency3.1sModel cost$0.015

Evaluation

passchecks 10/10 verdict computed from the gating checks

StatusKindCheckDetail
passstepenvironmentEnvironment check (snapshot, engine, content, key)
passsteplaunchLaunch run-match as a detached process
passstepsuperviseSupervise in the foreground until run-match exits
passsteprenderRender the 1080p recording
passstepnotesAppend Notes to summary.md
passcode_checkoutputs-existall present
passcode_checkgame-has-winnersonnet5 won: base destroyed
passcode_checkdecisions-loggedturns per slot: {'Multi0': 44, 'Multi1': 44}
passcode_checkllm-turns-validfailed/total per slot: {'Multi0': '0/44', 'Multi1': '0/44'}
passcode_checkvideo-decodesduration 60.4s
passcode_checkvideo-completevideo ends at tick 8760 of 8760
passcode_checklimits-respectedended after 9.1 wall-minutes (budget 30.0 min): base destroyed
passchecklistnotes-explain-winNotes section explains Sonnet5 victory via infantry assault and economic dominance, citing final thoughts from both models
passchecklistplayers-match-requestresult.json shows Multi0=sonnet5 (anthropic/claude-sonnet-5) and Multi1=gpt6luna (openai/gpt-6-luna) matching the requested players
passchecklistno-silent-outagerun.log shows continuous decision turns every 200 ticks with no gaps, and llm_failures=0 for both players across 44 turns each

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.

Ready to stop guessing if your outputs are good?

Create your eval Book a 15-minute walkthrough
Connect your agent to Jetty →