Saurav Panda 070ed5a8f2 Add report comparison tool and a heavy-DOM eval task ...

compare.py prints latency deltas with the pass rate in the adjacent column,
because a profile that is 40% faster and 30% less accurate is a regression
wearing a costume. Per-task rows sit below the summary so one pathological
task is visible rather than averaged away.

bulk_extract targets a 400-row, 1200-element page. The three original hermetic
tasks saturate at 100% pass and exactly 3 steps on every profile, which makes
them blind to step-count and accuracy effects; this one at least moves DOM
serialization from 0.12s to 0.76s per observation, making that term measurable.

2026-08-22 14:00:58 -07:00

..

hermetic

Add report comparison tool and a heavy-DOM eval task

2026-08-22 14:00:58 -07:00

web

Add dockerized eval harness with local and cloud browser backends

2026-08-22 11:19:30 -07:00