SemIf
Semantic ifs from open models, on a 3090 at home.
Same frozen 4B model. Same state. Same 21 questions. separate measured runs · aligned at t=0
21 DISTRIBUTIONS0 OUTPUT TOKENS
$ semif-score --mode shared …
reading shared state…

          
median 1.023s · 3 runs21 typed probability pairs

        


        

          
1 ORDERED JSON ARRAY111 OUTPUT TOKENS

          
$ python benchmarks/decision_vs_generation.py …
prefilling… waiting for first token…

          

          
median 5.332s · 3 runsvalid JSON · 21 values

        

      


      

        
Direct readout1.023 s

        
Compact array5.332 s

        
Wall-time ratio5.21×

        
Avoided output111 tok

      


      

        
        
Median of three warmed, sequentially measured runs on one RTX 3090 with the same frozen BF16 Qwen3.5-4B and one owned 21-criterion state. The generative prompt requests only an ordered yes/no array—no keys, confidence objects, or explanation. Commands are abbreviated; complete commands and raw token timelines are linked from the repository. The replay aligns both paths at t=0 and preserves their measured completion ratio.