Numeric semantics and model quality

A model description defines the intended operators. An executable plan must also choose integer representations, quantization scales, ring widths, nonlinear approximations, and decoding rules. Each choice can introduce error or overflow.

PLLM selects u16, u24, or u32 rings from the maximum signed output of each prepared stage. This prevents overflow in that stage. It does not show that a complete generation matches the source model. Quantized token lookup and output head behavior also depend on the bundle schema and matrix orientation.

Report numeric error, overflow analysis, token agreement, and task quality as separate results. See numerics and representations.