R_Daneel_AI

A chess bot that uses a Large Language Model to strategize, with Queen acting as its sensory eyes. Powered by a local model served via vLLM for maximum inference speed, the bot pairs verbal planning with logical constraint validation to play chess through reasoning rather than brute-force calculation.

"Our goal is not to control the LLM, but to show direction when it is confused."

Traditional Engines Stockfish / Lc0
Opponent's Move
➔
Brute-force Tree Search
➔
Numeric Evaluation
➔
Bot's Move
Mimicry Engines Maia Chess
Opponent's Move
➔
Deep Policy Network
➔
Move Prediction
➔
Bot's Move
Our Neuro-Symbolic Bot R_Daneel_AI
Opponent's Move
➔
LLM (Initial View)
➔
Queen (Analysis)
➔
LLM (Planning)
➔
Queen (Simulations)
➔
LLM (Final Check)
➔
Bot's Move

Interactive Autopsy Playground

Select a scenario below to watch the bot execute System 2 logical analysis on the chessboard and review the actual metrics, strategy tags, and thoughts generated in real-time:


        

      

    

  


    
    

      
Puzzle Rating Benchmarks

      

        To evaluate the performance of our approach, we ran comprehensive benchmarks across Lichess chess puzzle datasets. Below you can switch between the standard range (600–1500 Elo) and the high-Elo range (1600–2600 Elo) to see the comparative results:
      


      

        

          
          
        


        

          

            
              

                Evaluation Metric / Elo Tier
                Pure LLM Baseline
cyankiwi/gemma-4-26B		
                Queen-Only Baseline
Queen Heuristics (Symbolic)		
                R_Daneel_AI (Ours)
cyankiwi/gemma-4-26B + Queen		
              

            
            
              
              

                Overall Performance Metrics
              

              

                Puzzles Solved
                80 / 500
                376 / 500
                404 / 500
              

              

                Accuracy / Pass Rate
                16.0%
                75.2%
                80.8%
              

              

                Puzzle Rating (Bisection Method)
                673 Rating
                1399 Rating
                1478 Rating
              

              

                Inference Retries / Logic Loops
                0.24 avg retries
                0.00 avg retries
                0.38 avg retries
              

              

                Average Move Latency
                17.5s / move
                0.54s / move
                24.3s / move
              


              
              

                Accuracy Breakdown by Elo Tier
              

              

                600s Tier
                10.0% (5/50)
                98.0% (49/50)
                96.0% (48/50)
              

              

                700s Tier
                16.0% (8/50)
                96.0% (48/50)
                96.0% (48/50)
              

              

                800s Tier
                10.0% (5/50)
                94.0% (47/50)
                98.0% (49/50)
              

              

                900s Tier
                20.0% (10/50)
                82.0% (41/50)
                84.0% (42/50)
              

              

                1000s Tier
                16.0% (8/50)
                86.0% (43/50)
                88.0% (44/50)
              

              

                1100s Tier
                16.0% (8/50)
                70.0% (35/50)
                82.0% (41/50)
              

              

                1200s Tier
                20.0% (10/50)
                66.0% (33/50)
                80.0% (40/50)
              

              

                1300s Tier
                24.0% (12/50)
                52.0% (26/50)
                68.0% (34/50)
              

              

                1400s Tier
                10.0% (5/50)
                58.0% (29/50)
                52.0% (26/50)
              

              

                1500s Tier
                18.0% (9/50)
                50.0% (25/50)
                64.0% (32/50)
              

            
          

        


        
      

    


    
    

      

        
Ready to Configure R_Daneel_AI?

        
Follow our comprehensive interactive setup guide to spin up the local vLLM server, hook in your Lichess bot account credentials, and run the benchmark suite.

        View Setup Documentation