02 — Run · Round 2 · Round robin · r2-rr-jev-gpt56terra-g1
Jev Router vs GPT-5.6 Terra
GPT-5.6 Terra destroyed the enemy base after 10.2 game-minutes · 17.0 min wall clock
LOSS ukraine · west
Jev Router
typesafe/jev-router
Value destroyed$6,500Value lost$17,150
Units killed / lost39 / 32Buildings killed / lost0 / 7
Peak army$2,100Orders issued100
Decision turns (failed)77 (49)Mean latency6.1sModel cost$0.107
WIN england · east
GPT-5.6 Terra
openai/gpt-5.6-terra
Value destroyed$17,250Value lost$6,600
Units killed / lost33 / 40Buildings killed / lost7 / 0
Peak army$4,200Orders issued85
Decision turns (failed)77 (0)Mean latency3.8sModel cost$0.587
Evaluation
failchecks 9/10 verdict computed from the gating checks
| Status | Kind | Check | Detail |
|---|---|---|---|
| pass | step | environment | Environment check (snapshot, engine, content, key) |
| pass | step | launch | Launch run-match as a detached process |
| pass | step | supervise | Supervise in the foreground until run-match exits |
| pass | step | render | Render the 1080p recording |
| pass | step | notes | Append Notes to summary.md |
| pass | code_check | outputs-exist | all present |
| pass | code_check | game-has-winner | gpt56terra won: base destroyed |
| pass | code_check | decisions-logged | turns per slot: {'Multi0': 77, 'Multi1': 77} |
| fail | code_check | llm-turns-valid | too many failed turns: {'Multi0': '29/77'} |
| pass | code_check | video-decodes | duration 103.7s |
| pass | code_check | video-complete | video ends at tick 15252 of 15252 |
| pass | code_check | limits-respected | ended after 17.0 wall-minutes (budget 30.0 min): base destroyed |
| pass | checklist | notes-explain-win | Notes section cites final thoughts of both Jev Router and GPT-5.6 Terra and explains the base-destruction win |
| pass | checklist | players-match-request | result.json confirms Multi0=jev (typesafe/jev-router) and Multi1=gpt56terra (openai/gpt-5.6-terra) exactly as requested |
| pass | checklist | no-silent-outage | all 29 Jev Router LLM failures (504 timeouts) were properly captured and flagged by the llm-turns-valid code check; no unreported silent outage |
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.