Token Uncertainty Spectrometer

Where did the model hesitate?

Each token of the model's answer is colored by how uncertain it was when it chose that word. Calm green = near-certain. Hot red = a genuine fork. Hover any token to see the ten alternatives it weighed. Runs entirely in your browser against any OpenAI-compatible endpoint.

🔒 base URL, model & key live only in this browser (localStorage) — requests go straight to your endpoint
Temperature ≠ 1 distorts the signal. Logprobs are temperature-scaled — at 0 they collapse to a point mass (everything green); above 1 the spread is artificially inflated. Use 1.00 to read the model's true uncertainty.
awaiting a question · submit to render the token field
Reading: Shannon entropy
certainuncertain
Hover a token above to inspect its top-10 alternatives · click to pin.