Belief Values Are Attention Weights Over Hypotheses

In Elume, a BeliefState is an immutable numeric snapshot. The values are not sentences. They are mass assigned across a hypothesis space: which interpretation should influence the next step most?

1. Adjust A BeliefState

Raw inputs are normalized so the prior vector sums to 1.0.

Plain English 

        

      

    


    

      

        
2. What The Values Mean

        
The vector position is just a slot. The label comes from the caller.

      

      

        

          
High deadline

          

            [0.70, 0.20, 0.10] means the current state should be interpreted
            mostly through time pressure. Downstream components receive a
            deadline-shaped signal.
          

          urgenttime pressure
        

        

          
High uncertainty

          

            [0.15, 0.20, 0.65] means the system has not settled. Curiosity and
            competition layers may treat information-gathering as more valuable.
          

          ambiguousneeds evidence
        

        

          
Balanced state

          

            [0.33, 0.33, 0.34] means no one hypothesis dominates. This is a
            diffuse or evenly split interpretive state.
          

          diffuseno strong winner
        

        

          
Important boundary

          

            Elume's BeliefState stores the numbers, timestamp, and optional
            evidence. Names like deadline or workload are semantic labels owned
            by the caller or visualization.
          

          core is numericlabels are external
        

      

    


    

      

        
3. How It Moves Through Elume

        
Belief values become a trajectory feature vector, then temporal context.

      

      

        

          
Named hypothesis space

          
deadline, workload, uncertainty. This name map lives outside BeliefState.

        

        

          
BeliefState.prior

          
A 1-D float32 vector that sums to 1.0 and is locked read-only.

        

        

          
BeliefEmbedder

          
Projects the prior into a fixed-width TrajectoryStep state vector.

        

        

          
LinOSS Encoder

          
Consumes time-ordered steps and compresses their movement into context.

        

      

      

        
        

      

    


    

      

        
4. Prior vs Evidence

        
In this model, evidence is an optional likelihood vector over the same slots.

      

      

        

          
Prior

          

            The system's current distribution before a new observation is integrated.
            In active-inference language here, it is the posterior from the previous step.
          

        

        

          
Evidence

          

            Optional observation likelihood over the same hypothesis slots. It must be
            a 1-D float32 vector, but the current BeliefState class does not perform
            Bayesian updating by itself.
          

        

      

      

        The class is a record, not an updater. It preserves a belief snapshot; other
        operations decide how to create or revise that snapshot.