Interactive theory figure · fixed-prototype attention model

Same next-answer risk, different error clocks

Hold the current score margin fixed and change the effective attention mass. Both histories start with the same error probability, but each additional correct output changes their margins at a different rate.

Current margin F History A mass History B mass Limiting margin beta Future answers H Decoder temperature

Current error probability, both histories

First error within H answers · A

First error within H answers · B

· History A· History B

Scroll horizontally to inspect the full plot.

Move over the curves to inspect an answer step.

Exact law for the stated binary, fixed-query/prototype attention head. No positions, residual path, MLP or layer normalization. This is a mathematical construction, not a fit to Qwen. At each step, the hazard is conditional on every preceding answer being correct; cumulative risk uses the product of their survival probabilities. Johnson and Huo's starting framework.