Same prompt, two models, different outputs. No tooling was actually showing me where they diverged.
Built tokenflame that gives entropy heatmaps, tokenizer diffs, divergence markers, token-by-token replay. One command, one HTML file.
pip install tokenflame