Three egocentric trajectories from Aria Everyday Activities, paired with the
milestone decompositions generated by gemini-3.6-flash. Click any
milestone bar or evidence timestamp to seek the video to that moment.
⚠
Pilot, n=3 — timestamps are not yet reliable.
2 of 3 annotations place milestones past the end of their source video. Those
regions are drawn hatched in red beyond the video ends marker, and their
seek controls are disabled because there is no footage to seek to. The model
reported needs_human_review: false and confidence ≥ 0.97 on both.
Inferred task
Milestone timeline
Independent evaluation
Two graders re-scored these milestones from the video alone — given the evaluator
question and criteria, but never the reference timestamps, statuses, or task summary.
Both are Gemini-family, so this is a cross-model check, not a cross-vendor one.
Milestones
Completion and failure criteria are written to be gradeable on a
different trajectory, not just this recording.