Dataset Statistics

For PTS datasets this counts events at all three scales (latent / token / sentence) and the causal links between them. Emitted Δ-probabilities and latent readout scores are charted separately: they are different quantities and are not comparable.

Explore Causal Reasoning Events

Filter events across all three scales. Latent meta-tokens are observational: they carry a readout score, not a probability delta, and are shown in purple rather than green/red.

Event Detail

Causal Event Graph

Latent meta-tokens (diamonds) → pivotal tokens (circles) → thought anchors (squares) → outcome (star). Edges come from the event links (precedes_event_ids / linked_event_ids / parent_event_id). Green = positive impact, red = negative impact, purple = latent (no measured valence). Node size reflects score. legacy datasets fall back to the original reasoning graph.

Embedding Space Visualization

t-SNE projection of sentence/token embeddings. Explore clusters and patterns.

Reasoning Timeline

One shared generation axis across four scales: latent meta-tokens, emitted pivotal tokens, thought-anchor sentences, and the resulting success probability. Latent events are placed by their offset from the emitted event they precede.

Latent Workspace Heatmap

Meta-token (or category) x generation position, colored by readout score — how strongly the lens surfaces that concept. This is not a probability delta.

Step-by-Step Reasoning Circuit

Try a PTS steering vector

Vectors found by pts discover: pivotal-token search on a behavior eval, validated against controls on held-out questions. Load one, then compare the model's answer with and without it. Both answers use the same seed, so they differ only by the vector.

Run it yourself, on MLX or PyTorch, or serve it with an OpenAI-compatible API: see the vector's model card and PTS on GitHub.