Too many gate alerts: reduce the interruption without hiding the misses

By Gate & Shift Editorial Desk · Planned week of 2026-12-14 · Documentation reviewed 2026-11-15

If the phone sounds twelve times while the night supervisor is dealing with one ordinary delivery, staff may stop trusting the alert route entirely. Save examples first. The aim is a usable watch with an honest record of what it misses — not a silent phone at any cost.

This is a proposed tuning worksheet. It provides no vendor nuisance-alert ranking and no promise that a particular setting will solve a site problem.

Sort the interruptions before changing the rule

Use three buckets: the target activity really happened; ordinary activity met a rule you no longer want; or an alert has no matching target activity in the permitted evidence. Unresolved cases go in a fourth. An unwanted but correctly triggered delivery alert is a rule-design problem, not automatically a detector false positive.

Write one plain sentence defining the event that deserves a call, then add its camera, time window and boundaries. NIST's activity-specific definitions show why an event label needs observable conditions M01. “Suspicious” without a defined observable action leaves every shift judging something different.

Keep a short event ledger

For each alert, keep its time, rule version, matched reference event if any, action taken and observation status. Decide how repeat messages and grouped incidents will be counted before comparing systems — a system's incident count need not equal the number of objects in view.

Our field method matches alerts to human-annotated reference events using a declared rule M03, and it records unmatched alerts per monitored camera-hour M02. That is an operating-burden measure, with a different denominator from a negative-window false-positive rate.

Change one decision at a time

Propose a narrower schedule, a more relevant zone or a duration condition only if the actual product supports it. Retain the old rule and explain the change in the shift log. Pick the comparison window in advance and keep the same view where possible.

Check a labeled set of wanted events as well as the noisy examples — a rule can cut interruptions by missing the thing you care about. If weather, lighting or traffic changes substantially between windows, record it; don't pin the whole difference on the setting.

Carry counts with the fraction

Take a hypothetical watch that produces 18 scorable alerts. Twelve match eligible events and six do not, so precision is 12/18, about 66.7%. If the same period contains fifteen reference events and twelve matched, recall is 12/15 = 80%. The two numbers answer different questions M04.

If those six unmatched alerts came from 24 monitored camera-hours, the burden measure is 6/24 = 0.25 unmatched alerts per monitored camera-hour. None of these calculations says how the next month will behave. Keep a zero denominator undefined and report observation gaps separately.

Agree the response threshold with the people answering

Ask what staff can realistically review and what events must never be quietly discarded. Write the pilot acceptance decision with both missed reference events and interruption burden visible. A rule that looks tidy in a chart can still be impractical at shift change.

The pilot sheet keeps those decisions together. For a disputed missing event, use the missed-vehicle guide before assuming the quieter rule is working well.

FAQ

Which vendor produces the fewest false alerts?

The supplied evidence has no common labeled camera test across these vendors. A ranking needs shared reference events, boundaries, thresholds and monitored-time counts.

Is 0.25 a false-positive rate?

Not as used here — it is unmatched alerts per monitored camera-hour. A negative-window false-positive rate would divide flagged eligible negative windows by all eligible negative windows. Label the unit so nobody mistakes one for the other M02.