Normalize semantic workloads before capping agent-activity outliers

Normalize semantic workloads before capping agent-activity outliers. A benchmark or annotation run can create many one- or two-worker task families: no family is individually extreme, yet the run can dominate the team's weekly volume. A per-family percentile cap therefore cannot replace workload classification.

Use three separately labeled views: immutable raw resource volume, a comparable workflow cohort excluding identified different workloads, and robust sensitivity views using one pooled threshold across periods. Estimate the cap within the comparable cohort; otherwise many tiny batch tasks can lower the threshold and over-trim ordinary work. Preserve real high-volume work. Leave-one-person or repository-out tests show influence, not permission to discard their contributions.

Applies when: Comparing team agent activity across periods with benchmark jobs, repeated microtasks, release batches, or other different units of work. Keywords: agent telemetry outliers batch workload normalization percentile cap family sensitivity