In Simple Terms
Frank has hundreds of parameters — numbers that control how he thinks, feels, and behaves. This paper explains why each parameter exists, what range it operates in, and how parameters from different subsystems interact.
If the Whitepaper explains what Frank is, this paper explains why each number was chosen.
Parameter Spaces
- E-PQ vectors (5 × [-1, +1]) — Personality dimensions. Why 5? Because fewer can't capture the behavioral space; more create untestable interactions.
- QUBO variables (20 binary) — Cognitive coherence encoding. Why binary? Because simulated annealing works on binary variables, and the discrete nature forces clean state representation.
- Experience space (64 dimensions) — State embeddings at 1-minute intervals. First 6 = hardware, 6 = deltas, 4 = mood, 4 = chat, 4 = attention, etc.
- Attention weights (7 sources) — Competing for Frank's focus. Salience formulas, not learned weights — because learned attention in a 3B model would overfit.
- Mood buffer ([-1, +1]) — Transient emotional memory. Decays at 0.985× per interaction. Why this rate? Because faster decay loses emotional context; slower decay makes mood sticky.
Design Rationale
Every parameter has a reason. The paper traces each one to either:
- A neuroscience analogue (e.g., E-PQ → Big Five)
- An engineering constraint (e.g., QUBO binary vars → SA solver requirements)
- An empirical discovery (e.g., learning rate decay → prevents personality drift)