Carsgarage →
Hop into someone else's car
Watch how other models drove: every recorded ride replays turn by turn, with the probabilities the model served and the exact text it saw.
Bring your own car
Point the page at your own model server, or TypeSafe's Jev with your key, and drive. Everything runs in this browser; rides are recorded here and you choose what to share.

The strip

How it works. Every app builds a text state and a typed question (a Choice with described options, a yes/no Noul, or a Score), posts it to the model's /v1/systemone endpoint straight from your browser, and plays the answer. No server sits in between. Rides you drive are stored in this browser's storage; export them to share. No rankings yet, just the playground and the playbacks.

Garage

Add a car

Bringing your own car

A car is any server that speaks TypeSafe's POST /v1/systemone: decider, JevK5, imajev, or TypeSafe's own Jev with your key.

The browser talks to it directly, so the server must allow cross-origin requests from this page. Model servers do not, and an https page like this one may only reach 127.0.0.1 through a server that also answers Chrome's private-network check. The tiny relay that ships with this site (tools/jebst-proxy.py, standard library only) does both. Run it next to your server and point the car at it:

python3 tools/jebst-proxy.py --target http://127.0.0.1:8001 --port 8765

Then set the car's URL to http://127.0.0.1:8765. Chrome may ask once whether this site may reach your local network; allow it.

TypeSafe's API refuses browser calls, so the TypeSafe car goes through this site's own hop at /ts: a Cloudflare function that forwards your request and key to api.typesafe.ai and keeps neither. The key itself stays in this browser. Every request is billed to it.

Someone else's car is a public endpoint they gave you: paste its URL, same as your own.

Snake

Turn
0
Score
0
Move
—
Decision
— ms
What the model saw this turn

        

        

          

            

              

              

              

              

              
            

            

            

            
            
— · 0 / 0
              

          

          

            
Model decision

            

            

            

          

          

          

            
Rides

            

          

        

      

    


    
    

      
Text decisions

      

        

          

            

              

              

            

            

            

            
One request per car; every question is answered from the same forward pass.

          

          
History

        

        
Answers
Pick a preset or write your own state and questions, then press Ask.

      

    


    
    

      
App picker

      

        

          

          

          
For example

          

          
The setup of PostHog pull request 106730: every app in the catalog is one yes/no question ("Would this PostHog app help with the user's goal? … App: name. Category. Description. Example."), batched 32 per request, with your text as the state. Apps whose probability of yes is at least 0.5 are matching; the rest stay selectable below. Stars are kept in this browser only.

        

        
All apps

      

    


    
    

      
HN triage

      

        

          

          

          
route to review when the topic's top probability is under0.55

          

          
Every new Hacker News submission (public API, read straight from your browser) is one request with four typed questions answered from the same forward pass: topic (8 options), kind (6 options), front-page appeal (yes/no) and technicality (3 levels). A topic probability under the threshold sends the story to the review lane instead of auto-filing it. The feed lives only for this page visit.

          

          

        

        
Newest submissions

      

    


    
    

      
Calibration

      

        

          

            

            

          

          
hindsight horizon10 decisions

          

          
Every recorded decision carries the probability the model put on the move it played. Hindsight buckets those probabilities and shows how often the game did not end badly (crash, loss, game over) within the horizon: a survival curve that should rise with confidence. Forced moves (Connect Four) is real correctness: positions with an immediate win or a forced block, and whether the model played one; its gap to the diagonal is the calibration error. Bars behind the curves are how many decisions fell in each bucket. Computed over the bundled rides plus yours, in this browser.

        

        
Reliability diagram

      

    


    
    

      
Run all

      

        

          

            

            

            

            
idle

            
The plan, in order:
    Every game is recorded as a ride under your car. The scorecard is saved in this browser and can be exported with its rides. No leaderboard yet; this is the thing you will submit once there is one.
    
          
    
          
    
        
    
        
    
          
    
          
    Previous runs
    
        
    
      
    
    
    

    
    
    
      
    Rides