Kinetic DNA rows
—
Active clusters
—
Warehouse sessions
—
Signal integrity
—
Behavioral discovery cloud

K-means clusters on kinetic fingerprints

Scatter uses PCA on all 16 embedding dimensions (same rows as the parallel plot below). Use Cloud points to plot one dot per kinetic row, per FullStory session (mean fingerprint), or per visitor (mean across sessions when nexus_user_key is present). Data comes from warehouse.jsonl via /summary.

What are PC1 / PC2, and when is “fallback” used?

The idea: Each row has a 16-number fingerprint (a vector). You can’t plot 16 axes at once on a flat screen, so PCA finds the two directions in that 16-D space where the data varies the most. Those become PC1 (horizontal) and PC2 (vertical). Each point’s position is its projection onto those two directions—so nearby points are “similar” in a variance sense across all dimensions, not just raw columns 0 and 1.

Fallback: If there aren’t enough distinct samples (or variance collapses), PCA isn’t reliable. The chart then falls back to plotting raw dim 0 vs dim 1 so you still see something—those are only two coordinates of the embedding, not the full picture.

New captures include challenge_module from the lab page. Older rows are matched by label prefix (e.g. SR_*, SPEED_*). Rows that don’t match are omitted or grouped under Other.

Session and visitor modes collapse rows before PCA. Visitor mode needs nexus_user_key on captured rows.

4

How to read it: Optionally narrow rows by challenge module (derived from labels) and choose whether each dot is a kinetic row, a session mean, or a visitor mean. K-means runs in the PCA plane on that subset. Colors identify clusters (same order as the swatches). Click a dot or a session to highlight paths and update the radar. When FullStory was active during capture, use Open in FullStory beside the radar to jump to that replay moment.

Session radar

Persona shape (this session)

Select a session

Choose a session from the list or click a point in the cloud to decode behavioral DNA from captured rows.

Organizational hook

This view shows how intent becomes kinetic when you stream behavioral DNA into a warehouse and cluster it. Want to find emergent archetypes on your own platform?

Full fingerprint shape

Parallel coordinates (16 dimensions)

One polyline per kinetic row; axes 0–15 are min–max scaled within the current warehouse so you can compare shape across samples.

Together with PCA: the scatter compresses 16-D into 2-D for clustering; this view preserves all dimensions so you can see which axes separate sessions. Selecting a session in the list emphasizes its rows.

Embedding dimensions

Per-dimension strip charts

Sixteen charts—one per fingerprint coordinate 0–15. Use session or visitor means (horizontal bars), or per kinetic row (scatter: one dot per row, color = FullStory session). Narrow by challenge module or pool Any module.

Independent from the behavioral cloud’s “Cloud points” control.

Only kinetic rows matching this module feed the means (or all modules when Any).

Discovery dashboard

Fingerprints, warehouse, and charts

A short guide to how kinetic fingerprints are produced, how they land in this view, and how to interpret what you see.

1 · From sessions to fingerprints

In the lab, FullStory captures interaction bundles (pointer moves, scrolls, etc.). A web worker normalizes those events into short motion features, runs them through a trained sequence model, and emits a 16-number vector (the fingerprint) when movement energy crosses a threshold and a challenge phase label is active.

Each capture is written to the collector as JSON (for example in warehouse.jsonl) with fields such as label, session_url, and fingerprint. This dashboard loads that warehouse via /summary—it does not recompute embeddings; it visualizes what was stored.

2 · How fingerprints feed these charts

  • Behavioral cloud — Uses kinetic rows only. Your filters choose which rows enter the analysis. Optionally rows are averaged per session or per visitor before clustering. PCA projects the 16-D vectors onto two axes (PC1 / PC2) so points can be plotted; k-means assigns cluster colors in that same plane.
  • Session radar — Summarizes the selected session using kinetic rows (aggregated signal shape), so you can compare “persona” intensity across a few interpretive dimensions.
  • Parallel coordinates — Draws the full 16 dimensions per kinetic row (min–max scaled in the current warehouse), so you see shape beyond what PCA compresses.
  • Dimension strips — One small chart per fingerprint coordinate (0–15). You can view session means, visitor means, or raw per-row values depending on the Units control.

3 · How to read the charts

  • Stats row — High-level counts: kinetic rows loaded, number of clusters requested (subject to how many points you have), distinct sessions, and a coarse integrity hint.
  • Cloud scatter — Distance in the plot is similarity in the PCA plane (variance directions), not raw finger distance in all 16 axes. Same color = same k-means cluster. Click a point to drive session selection and the radar.
  • PC1 / PC2 caption — If you see a “fallback” note, the view is plotting two raw embedding dimensions instead of PCA—usable for demos with few points, but interpret cautiously.
  • Radar — Read as a profile for the chosen session, not a universal scorecard; it encodes how kinetic signals load into the archetype axes used here.
  • Parallel + strips — Use these when you care about specific embedding dimensions or comparing sessions side by side along one axis at a time.