Replay of the AdSERP dataset with per-AOI four-class taxonomy labels overlaid: CLICKED DEFERRED EVALUATED-REJECTED NOT-APPROACHED. Labels derived from cursor enter/dwell/exit episodes against AOI bboxes — the inference task the AR library performs in production.
Curation rubric: docs/CURATION.md · Sibling: live testbed · Pool: 31 grouped + 110 extended = 141 total trials
3+ active label classes in one SERP — comparative inference
cursor entered, retreated, returned
approached and dropped, no re-approach
dwell near the 100ms classification threshold
gaze-fixated AOIs that cursor missed entirely
one click, no other approached AOIs
Orthogonal cut by SERP layout — bbox kinds present, not behaviour. Use this to spot-check whether AOI bbox extraction holds across composition types. Population share is across all 2776 AdSERP trials, not just the audit pool.
High-scoring trials not yet assigned to a pedagogical group. Browse to spot-check the curation or surface trials worth promoting.
attentional-foraging/scripts/extract_organic_bboxes.py). AOI labels from derive_aoi_labels() in scripts/build_replay_trial.py.