Recipes, not bare models
A recipe is a model plus how it is asked: prompt style, the vault context it gets, and its tools. Arena tests the whole recipe.
Ghost replays of your own tasks
The prompts you actually gave your AI tools, replayed on two recipes at a time. Prompts that only made sense in their old conversation are left out.
Leagues and weight classes
Every domain and kind of task (drafting, decisions, research, planning, video scripts, coding) is its own league, split into light, middle and heavy by difficulty.
Checkers first
A number that must reconcile, valid JSON, code that passes its tests, a word limit, citations. When a checker can settle a duel, nobody has to.
Two-second blind picks
Only where the answers really differ: two answers, names hidden, one key to pick. Then the names are shown, with the unblinding score: how often your blind pick matched the current route.
Equivalent, use the cheaper
When two recipes come out even, the cheaper one gets the route, and the policy says how much cheaper.
A routing policy with reasons
Plain files in your vault. Glyph Studio's Auto follows it and says "routed by Arena"; every route carries one sentence on why.
Home Only league
Health and money run on local models only, on this Mac or your own machines, refused before any call. Nothing is sent.
Adverse-event register
Flag an answer that made something up or gave unsafe advice. One severe event, or two major ones in 30 days, suspends that model in that domain.
Spend cap and sentinels
A monthly cap is checked before every cloud call. Six fixed tasks run weekly to catch a model that quietly got worse.
Scouting
New local models and the overnight web scout's picks show up as challengers, one click to add.
Open source
MIT. Arena's logic lives in Glyph Engine, next to the model providers it uses.
