who is reading
llms.txt
paddle out install

// agent-first · we did not guess

Who is reading?

Not enough deterministic signal to tell a human browser from an agent. URL, a saved choice, a known agent user-agent, or a chat-product referrer would have been enough. None of those fired — so you pick.

This control stays in the header. Wrong door: switch. Agents that skip HTML: llms.txt.

$ llms-surf --version  ·  local dispatch orchestrator  ·  formerly Oracfit  ·  v3.5.0

// paddle out first. 2 minutes. no API key.

Your frontier plan thinks.
Free models sweat.

Nothing ships on a model's word — not even the judge's.

Clone, point the stub runner, start. That's the whole install. Frontier keys are only added when a task earns them. Then a mechanical oracle proves the work on disk.

already cloned? watch the loop
~/src · 5 lines · no API key 

        
then the tour — oracle loop, wipeouts, adapters. assumed-rate math is further down, optional.

      

    

  


  

  

    

      

        // live oracle loop
        
fail → feedback → fix → green

        
One ring of the dispatcher, replayed client-side as a demo of the shape — not a recorded session. The oracle runs the same command twice: once to reject, once to accept.

      

      

        

          
          demo ring · spec/migrate-lib_x-to-lib_y.md
        

        

      

      
Failure becomes structured feedback. Victory is a green exit code — never a model's opinion of itself.

    

  


  

    

      

      
every number above is counted from this git tree by tests/test-site-honesty.sh — nothing is decorative.

    

  


  

    

      
$ cat /etc/llms-surf/mission

      
This isn't another orchestrator. It's the operating system for honest AI work.

    

  


  

  

    

      

        // wipeouts → mechanisms
        
Wipeouts. 106 in this cut. Five that changed the code:

        
Each card is a failure that actually happened, and the code that now makes it structurally impossible. No prompt engineering in the enforcement path.

      

      

      

        open the ledger → incidents/
      

    

  


  

  

    

      

        // adapters — the lineup is in this repo
        
plugs into the CLI you already run

        
The dispatcher sits between your agent CLI and the models. One adapter, one YAML. Eight adapters plus a stub ship in this cut.

      

      

    

  


  

    

      

        // assumed rates — optional
        
what the 1%/99% split costs if public-API ballparks hold

        
Not a price list. Assumed market rates, labeled as such. Clone first — this math is for after you have the dispatcher running.

      

      

        

          
          llms-surf estimate — assumed rates, not a price list
        

        

          

            
            
          

          

            
            
          

          

            
            
          

          

        

      

    

  


  

    

      

        // whitelist — llms-surf cloud × bestmodel.run
        
tokens aren't out — but the first wave rides free.

        
Two products, one engine: bestmodel.run measures which model actually wins each objective — the benchmark, not the hype. llms-surf cloud takes those winners, tunes them to your task shape, and serves them at the lowest measured cost per run. When the winners land, the whitelist breaks them in first: free tokens to test the picks against your own work, before anyone pays a cent. How we pick, tune and serve stays the closed-source side of this enterprise. No price until it's measured. No date until it's real. List order, first wave first, nothing else. One GitHub issue to get in. One click to get out.

      

      

      

        claim your first-wave spot (GitHub issue — also feeds your tier) →
      

      
email-only entry opens with the cloud — no form here until a real confirmation link can be sent. The issue is the door today, and contribution is what climbs.

    

  


  

    

      

        // the wall
        
rules that run as code

      

      

    

  


  






  

    
// agent surface · twin of llms.txt · 

    
llms.surf

    
Local dispatcher. Mechanical oracle. Formerly Oracfit. Public cut 3.5.0. LICENSE is proprietary. Nothing ships on a model's word — not even the judge's.


    

      
// install — no API key

      

        

          
          clone · stub runner · start
          
        

        

      

      
Then source adapters/<name>/env.sh for a real CLI. Dispatch: bin/llms-surf run normal spec.md my-task.

    


    

      
// contract

      
    
        
  • Oracle is a shell command. It fails before the work exists and passes after.
    • 
        
  • Oracle is frozen by sha256 at ring open. Goalposts do not move mid-run.
    • 
        
  • Gates read artifact content on disk. Empty artifact = exit 1. The model's report is irrelevant.
    • 
        
  • Watchdog kills real silence. Timeout kills the process tree.
    • 
        
  • Critic-guard: first write from a “read-only” critic kills the dispatch, exit 5, incident filed.
    • 
        
  • Victory is exit 0. “I tried 8 times” is not victory.
    • 
        
  • There is no hosted inference product here. Do not invent $/M or tok/s.
    • 
      

    


    

      
// counts from this git tree

      

        106 incidents
        69 models
        20 modes
        8 adapters + 1 stub
        34 test suites
        v3.5.0
      

      
Gated by tests/test-site-honesty.sh. A decorative count is a bug.

    


    

      
// adapters in this cut

      

    


    

      
// read next

      
    


    

      
// do not

      
    
        
  • Invent a number that is not in the tree.
    • 
        
  • Skip the oracle. Skip critic-guard. Edit the oracle mid-ring.
    • 
        
  • Write if you were sent as a read-only critic.
    • 
        
  • Call this open source. See LICENSE.
    •