G
Trajectory Grammar Atlas · conversations in the latent space of a grammar

Every conversation is a walk through a grammar

You are looking at the latent space of a trajectory grammar: 7 progress states (greet, auth, intent, tool, clarify, confirm, close). Each of the … conversations below — the official, publicly released tau-bench evaluation runs (gpt-4o and claude-3.5-sonnet, airline + retail) — is a walk through those states, and the map is the PCA projection of each walk’s state-visit features. The grammar is not drawn by hand: … transitions collapse into … distinct edges, and the six heaviest carry … of all movement.

What to do: click a table row or a map point to read one conversation’s walk and transcript. Drag a box on the map to pick a region of the latent space: the machine below is re-induced from exactly the conversations living there. Filter by domain or model to compare the grammar of different regions of the space.

Domain
Model
Color by
Conversations click a row · click turns / D to sort
ConversationFirst user message Turns D
Latent space PCA of grammar-walk features · drag a box to select a region
The induced grammar edges = empirical state-to-state transitions of the selection, with counts

Flow rules (most frequent edges)

Backward and self edges (the loops)

Most common walks (flow templates)