Phase 4 of the tiny-adder experiment — real Ollama output, greedy decoding

Showing Its Working, Token by Token

The scratchpad model answering 472+38 through Ollama. Every branch is a token that received nonzero probability at that step; the copper path is what greedy decoding actually chose. Ollama routes the <think> and </think> tokens to the thinking field, so they are absent here; everything between them is the model's working.

prompt

token span Ollama's API labels message.reasoning

This is a computer process, not thought: each step is one forward pass producing 20 numbers (one per token in the vocabulary), and greedy decoding picks the largest. The near-certainty on most steps is what a fully learned algorithm looks like from the outside.

sampled — was also the top-probability token sampled — fell outside the shown top-k despite lower probability nonzero probability, not sampled

Built with vis-network from a real trace via Ollama's logprobs API: at every step the model's top candidate tokens and their real probabilities were captured, and the next token drawn by the server's own sampling -- not scripted.