TaskAudit-faithful three-stage pipeline · IMDb · GPT-5.6 Luna xhigh
Loading 30 recorded Stage-2 accessibility detections…
All Stage-2 accessibility positives are retained. Stage-3 taxonomy results are shown separately as MAPPED, UNMAPPED, or OUT_OF_SCOPE and never revise the Stage-2 decision. Counts are raw action-level findings, not deduplicated website bugs.
Before/after screenshots and focus rectangles come directly from the original agent runs; no browser replay was used. Red boxes show the recorded focus for each action.