Action rate variance
Measured on Pareto's own data source

demos of one task. The action rate spans .

Pareto ships per-episode reductions only. Its velocity-debiasing direction depends on a cross-episode number, and that number isn't shipped. So I computed it, on real public data, with no model in the loop.

Demos
Action rate span
min to max
Coefficient of variation
across demos
Rate outliers
5% tails
The measurement

Every demo of the same task, sorted by action rate

Each bar is one demonstration. Red bars are the 5% tails. The dashed line is the task mean.

middle 90%
5% tails (20 demos)
Shipped vs. missing

What Pareto exposes, and what it doesn't

Shipped measure_episode_trajectory

    All four are computed inside one episode. None aggregates across demonstrations.

    Missing: cross-episode action rate

      The same signal peak_velocity reads, aggregated across the demos of a task. This is the input velocity debiasing needs.

      What it surfaces

      The 5% tails

      Twenty demos whose speed, not skill, drives the variance. These are the prune candidates a cross-episode metric would flag.

      High rate
      Low rate

      Computed from lerobot/pusht, 25,650 frames across 206 demos. No model in the loop. Reproduce with node local_test.mjs.