A Real Model, Pinned

The practical phase starts with a real TinyStories run. The environment, prompt, greedy setting, image digest, and trace hash make the run reproducible by re-execution.

highlighted = computed this step

This book shifts from exact toy engines to a real TinyStories model run. The Docker image, model revision, seed, prompt, greedy setting, and trace hash pin the computation for re-execution.

\text{trace hash}\rightarrow\text{re-execution check}

The captured run

The fixed prompt is Once upon a time. The captured output text is Once upon a time, there was.

\text{prompt}\rightarrow\text{captured output text}

Summary

The discrete spine remains exact. Continuous magnitudes now enter the REAL CAPTURED register: literal floating-point cells, shown with approximate labels and backed by raw tensor bytes.

\text{exact IDs}+\text{captured floating-point magnitudes}