AT-301g · Module 1

Establishing Metric Baselines

3 min read

An alert without a baseline is noise. Baselines establish what "normal" looks like so that deviations are meaningful. The baselining process runs over two weeks of steady-state operations — long enough to capture daily and weekly patterns, short enough to not include pre-stability chaos.

For each metric, establish: the mean, the standard deviation, the daily pattern (is there a morning spike? a Friday dip?), and the acceptable deviation range. CLOSER's average cycle time is 47.3 minutes with a standard deviation of 12.1 minutes. A cycle time of 72 minutes is 2.04 standard deviations above the mean — worth investigating. A cycle time of 59 minutes is within normal variance — not actionable.

Baselines must be recalibrated after any structural change: new agent deployment, role redesign, process modification, or team scaling event. The baselines from a 17-agent team are invalid for a 20-agent team — coordination overhead changes the system behavior at every layer.

Do This

  • Baseline over two weeks of steady-state operations — enough to capture patterns
  • Track mean and standard deviation — anomalies are deviations, not absolutes
  • Recalibrate baselines after structural changes — new agents invalidate old baselines

Avoid This

  • Set arbitrary thresholds — "latency over 100ms" means nothing without context
  • Use day-one metrics as baselines — early operations are unstable by definition
  • Keep baselines static for months — the system evolves, baselines must evolve with it