Mine driving maneuvers into reviewable scenes

01 · Role in the system

What Jev does here

Adapters convert motion, lane geometry and object interactions into one readable evidence schema, then each sample gets lateral and longitudinal decision questions in a single Jev request with probabilities per answer; deterministic code merges consecutive speed phases into complete scenes (turn right, then lane change, then keep lane) with timestamps. Evidence stays inspectable: the review UI synchronizes camera footage, bird's-eye view, a timeline, raw probabilities and ground-truth editing, and missing evidence stays unknown instead of guessed. A recorded Boston nuPlan run with Jev 1.13 classified 933 fresh samples in 25.9 seconds (8 workers) for an estimated $0.32 - about $0.00034 per sample - with no local inference GPU.

02 · Control boundary

Where Jev sits

Jev as the per-sample labeling layer of a scenario-mining pipeline: typed lateral/longitudinal choices over a shared evidence schema (both questions share one request), code owns evidence assembly, scene merging and timestamps, and everything lands in a review UI with raw probabilities and editable ground truth.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• Performance numbers are one author-recorded Boston run; cost varies with context length and question count.
• Review footage and evidence excerpts are vehicle sensor data that leaves the machine to the API.
• Formerly named JevSceneMiner; the Python package and CLI keep the old name for compatibility.
• Decision labels are plain-language config (configs/labels.yaml); recall against human labels is not benchmarked.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• bskkimm ↗Community · github · public · checked 2026-10-15