size = the model's own confidence · the gold thread = the order it walked · scroll to zoom, drag to movesize = confidence · pinch to zoom · drag to move · tap a word
Ask anything. Then watch what it almost said.
Every answer a model gives you is one path chosen out of thousands it was weighing. This instrument draws that choice live, from the model's own numbers: the words it said, the words it set aside, and the private thinking in between.
In 2025, Anthropic researchers went looking inside Claude for patterns of neural activity that the model could put into words. They found them, and named the collection the J-space, after the Jacobian, the mathematical tool that surfaced it. Then came the striking part: given a multiplication to solve in silence, Claude's J-space lit up with the intermediate numbers, one after another, before any answer appeared. Words the model was thinking with and never saying.
The true J-space lives in a model's internal activations, and reading it takes the weights and a research lab. So this observatory does the honest next thing: it rebuilds the closest public shadow of that inner space from real signals the model reports about itself while it writes to you.
Gold, the words it said. Each one is placed by meaning and sized by the exact probability the model assigned it. Calm, bright words were near-certain. Small, flickering ones were close calls.
Turquoise, the words it weighed and set aside. At every step the model holds probabilities over thousands of possible next words. The strongest rivals surface here, joined to the fork by a fine turquoise thread: the sentence that could have happened, branching away from the one that did. Hover any gold word to light up the paths it beat.
Cream, the inner voice. Some models think in private text before they answer. When yours does, the ideas inside that hidden stream appear here first, the way 21, then 42, then 49 appeared inside Claude.
The geometry is real too. Every word is turned into a meaning vector of 2,048 numbers by a separate embedding model, sketched down to 24 dimensions with a seeded Johnson-Lindenstrauss projection (a 1984 theorem showing you can compress high-dimensional geometry while keeping its distances honest), then laid onto the two directions that spread this conversation's words the widest. Words that mean similar things sit near each other. The gold thread traces the order the model walked through them.
The strip underneath is uncertainty, word by word, measured in bits from the model's own probabilities. Tall bars are forks: moments the answer genuinely could have gone another way. Drag the timeline and you can replay the whole answer being decided.
And what this instrument cannot see, named plainly: attention, activations, and the true J-space itself. Those need the model's internals. For that story, watch the research film.
Your conversation goes to the model and comes back. Nothing is stored here. Uncertainty is estimated from the five strongest candidates the model reports at each step. A J-Space instrument built by Nicola Harvey, for Her Frontier.