Kokos

🇯🇵 JP Model Bench

Compare aisubscription (GLM pool) vs fugu-ultra (Sakana) on frozen JP fixtures. Runs dispatch the real refresh pipeline via Horizon.

Fixture

Models (comma-separated handles)

Use aisubscription for GLM pool, fugu-ultra for Sakana.

Capabilities

Run name (optional)

Generate narrative diff (A/B prose comparison)

jp-1 jp-10 jp-2 jp-3 jp-4 jp-5 jp-6 jp-7 jp-8 jp-9

Name	Status	Winner	Cost
jp1-smoke-6mdl	running	—	$0.0000
jpfugu2	completed	—	$0.0000	View
jpval-fugu	completed	—	$0.0000	View
jpval-kimi	completed	kimi-k2.7-code	$0.2358	View
jpval-glm	completed	aisubscription	$0.0284	View
jpval-deepseek	completed	deepseek-v4-pro	$0.0901	View
jp-cmp-kimi	running	—	$0.0000
jp-cmp-deepseek	completed	deepseek-v4-pro	$1.0028	View
jp-prod-glm	completed	aisubscription	$0.3648	View
jp-full-glm	running	—	$0.0000
jp-full-kimi	completed	—	$0.0000	View
jp-full-gemini	completed	—	$0.0000	View
jp-full-deepseek	completed	—	$0.0000	View
jp-smoke3-3	completed	aisubscription	$0.0470	View
jp-smoke2-3	completed	—	$0.0000	View
jp-smoke-3	completed	—	$0.0000	View
eval-topic-rerun-clean	completed	aisubscription	$0.0020	View