30-Day Software Delivery Load-Test Scorecard

Compare one non-critical repository across a baseline period and a controlled agent-assisted period. Keep the team, work intake, definition of done, CI configuration, review policy, and release process comparable. Record every material change.

Scope and controls

Repository

Team and cohort

Baseline dates

Assisted dates

Unchanged definition of done and target environment

Agent seats, eligible tasks, and unchanged controls

Measures

Measure	Fixed definition	Baseline	Assisted	Interpretation and context
Delivery outcome Delivery throughput	Items meeting the unchanged definition of done and reaching the target environment
Activity Concurrent branches	Daily median and peak active branches using the predeclared activity window
Leading indicator Push latency	Median and p95 from push start to server acknowledgement
Leading indicator Fetch latency	Median and p95 from fetch start to usable completion
Leading indicator CI queue time	Median and p95 from job eligibility to runner start
Context CI execution time	Median and p95 from runner start to job completion
Leading indicator Merge conflicts	Manual resolutions, total and per merged change
Review burden Reviewer minutes	Minutes reviewing, clarifying, requesting changes, and re-reviewing
Quality outcome Escaped defects	Confirmed period defects found after the agreed pre-release quality gate
Quality outcome Rollback frequency	Reversed, disabled, or emergency-hot-fixed releases divided by releases

Context to retain: work-item size or class, releases, CI execution time, outages, absences, configuration changes, and missing observations. Do not compare raw counts when the periods contain materially different work.

Twice-weekly review and stop conditions

Predeclared stop conditions

Material changes and confounders

Day 30 decision Scale: throughput improved and review burden, escaped defects, and rollback frequency stayed within predeclared local tolerances. Pause: apparent speed was offset by burden, defects, rollbacks, or an unsustainable queue. Investigate: the result was mixed, confounded, or too small to support a decision. Evidence and rationale