Skip to content
finalstate.aiThe Base Rate Trap: When the False Alarm Rate Decides
115Atlas

115THE WATCHMAN'S ALARM

The Base Rate Trap: When the False Alarm Rate Decides

A rare event can leave most alerts false even when a detector rejects 99% of non-events. Choosing a usable threshold also requires recall, investigation capacity and the costs of both errors.

An illustrative watchman faces a tray of unresolved alert records.Several schematic alert slips collect beside the watchman. Their truth is unknown until checked. The illustration supplies neither a measured queue size nor an automatic decision to mute alerts.WatchmanUnresolved
Illustrative watch. Alert slips are unresolved records; their number is schematic.

RARITY MOVES THE GOALPOSTS

Start with the population behind the alerts.

One event and 9,999 non-events form separate denominators in an illustrative 10,000-case population.An event marker sits beside a stack representing all 9,999 non-events. Labels state the counts; the objects are not an area chart or one mark per case. A 1% false-positive rate acts only on the non-event group.10,000 cases1 event9,999Non-events1% false positives
Illustrative counts. The single event marker and aggregate stack are not drawn in proportion to population size.
  • Recall counts detected events among all events.
  • The false-positive rate counts false alerts among non-events; specificity is one minus that rate.

In an illustrative population of 10,000 cases, one is an event and 9,999 are non-events. A 1% false-positive rate applies to the 9,999. Precision instead counts true alerts among all alerts. These denominators answer different questions.

NINETY-NINE OF A HUNDRED

One true hit, about 100 false alerts.

A rounded illustration contains one solid true-alert circle and 100 outlined false-alert squares.The fixed 101 marks round the expected 100.99 alerts to one true hit and 100 false alerts. The precise calculation alongside the figure assumes one event among 10,000 cases, perfect recall and a 1% false-positive rate, giving precision of approximately 0.9902%. Motion traces which denominator is used without changing the plotted counts.1 true≈100 false1 of ≈101 alerts
Rounded illustration: one solid circle is a true alert; 100 outline squares represent about 100 expected false alerts. Assumptions: one event in 10,000, 100% recall, 1% false-positive rate.

Assume perfect recall and a 1% false-positive rate. Expected false alerts: 9,999 × 0.01 = 99.99. With one true hit, expected total alerts are 100.99. Precision is 1 ÷ 100.99 ≈ 0.9902%; about 99.01% of alerts are false. These are expectations under the stated rates, not observed counts. Low precision creates checking work; its acceptability also depends on the cost of misses and the value of a true hit.

THE PERSONALITY SKETCH

The stated prior changed more than the judgments.

Stated engineer shares of 30% and 70% are compared with mean estimates of 50% and 55% after the same five descriptions.The fixed bars use a common zero-to-one-hundred-percent scale. Solid black bars denote the low-prior group; outlined accent bars denote the high-prior group. Estimates average each participant's mean over five descriptions; group sizes were 85 and 86. The null-description responses, not plotted, matched 30% and 70%.Engineer (%)Low priorHigh priorStated share30705055Mean estimate0100
Kahneman & Tversky, On the Psychology of Prediction (1973), reprinted chapter pp. 53–55, Prior versus individuating evidence. Low-prior n=85; high-prior n=86.

Kahneman and Tversky gave two groups the same five personality descriptions, with different stated mixes of engineers and lawyers. Average estimates of the chance of being an engineer were 50% in the 30%-engineer condition and 55% in the 70%-engineer condition. The prior effect was small but statistically significant. Without a description, judgments matched the stated priors. The finding is conditional, not evidence that people always ignore base rates. Read the original chapter.

INSENSITIVITY TO PRIOR PROBABILITY

Insensitivity to prior probability of outcomes.

Tversky & Kahneman, Judgment under Uncertainty: Heuristics and Biases, 1974

The section heading in their 1974 Science paper, pp. 1124–1125, names a judgment error. The mathematical prevalence effect remains even when everyone understands it. A fluent AI explanation still needs evidence about what the flag predicts.

THE INTRUSION DETECTOR

An alert queue consumes investigation time.

Unresolved alert records wait beside a single investigation desk.A schematic stack reaches a review position, where one unresolved record is examined. The remaining records wait. No numerical capacity, measured fatigue threshold or automatic muting is implied.WaitingInvestigate
Schematic queue. Axelsson (2000), §§5.1–5.3, pp. 191–194; Vaughan (2006), NASA Revisited, pp. 358–359. The detector model and organizational analogy are distinct.

Axelsson’s intrusion-detection analysis, §5, derives the constraint from a hypothetical installation and stated prevalence assumptions. It does not measure a universal fatigue threshold. Vaughan’s normalization of deviance describes organizational acceptance of anomalies at NASA. Applying that idea to a team that routinely dismisses alerts is a separate analogy, not an outcome established for this detector. Vaughan’s account.

A quieter queue can also miss more events.

A higher threshold selects fewer cases for investigation and can lose true hits. Compare operating points at deployed prevalence using false-positive cost, miss cost, required recall and investigation capacity. Precision alone cannot choose the threshold. Axelsson, §6.1, explains the detection and false-alarm tradeoff.

WHAT SURVIVES THE FLOOD

Agreement filters both classes.

An AND filter excludes a true event seen by only signal A while retaining a shared false alert and a shared true alert for an unresolved probe.Three schematic cases illustrate possible outcomes, not empirical frequencies. Solid circles mean true events; outline squares mean non-events. A true event flagged only by A is missed by the AND rule. A false alert and a true alert seen by both signals reach a probe whose result remains pending. The image assumes no independence and shows no automatic success.ABANDMissedProbePending
Schematic cases, not rates: solid circles are events; outline squares are non-events. A second requirement can exclude a true hit while preserving a shared false alert.
  • Measure the second signal among cases already flagged by the first, separately for events and non-events.
  • Count the true events excluded by the combined rule.
  • Route retained alerts to a defined probe; record failures and unresolved results.

An AND rule cannot increase either class's alert rate when the original flags are fixed. It may remove false alerts, lose true hits, or retain the same correlated errors. Multiplying marginal error rates requires conditional independence within each class; different-looking channels do not establish it. Judge the joint rates before sending retained cases to a real-world verifier.

THE FIRST ENEMY

Measure what the alerts mean where they will run.

An unfinished deployment ledger leaves true-alert and total-alert counts open for review.The ledger has blank entries for true alerts and all alerts. Its fraction remains unresolved because the illustrative calculation does not estimate deployment performance. The open review bracket confers neither approval nor a validated abstention policy.True alertsAll alertsReview
Deployment diagnostic. Fill the counts from a representative evaluation, with outcome follow-up; also measure missed events.

Estimate the share of alerts that are true in the deployment population, and measure what the detector misses. At the jagged edge, a fluent explanation cannot supply that validation. An abstention policy also needs testing: who receives unresolved cases, how often useful answers are withheld, and what errors still pass?

Read the transcript

01 · THE WATCHMAN'S ALARM

For an illustrative calculation, give a watchman's detector perfect recall and ninety-nine percent specificity. Those specifications describe its behavior within two different groups. They do not tell the watchman how many arriving alerts will be true. Alerts arrive in the same tray, each asking for an investigation. If the event is rare enough, most can be false. The watchman still needs to know what the detector misses, what a mistake costs, and how much checking the queue will require. Muting the alarm is one possible response, not an inevitable ending to this story.

02 · RARITY MOVES THE GOALPOSTS

Start with ten thousand cases. One is an event; the other nine thousand nine hundred and ninety-nine are non-events. Recall uses the event group as its denominator: of the events that occurred, how many did the detector catch? Specificity uses the non-event group: how many quiet cases stayed quiet? Ninety-nine percent specificity means a one percent false-positive rate. That one percent applies to the large non-event group. Precision uses a third denominator, the alerts themselves. It asks what share of those alerts are true. A small rate applied to a large group can produce far more false alerts than the rare events produce true ones.

03 · NINETY-NINE OF A HUNDRED

Keep the same population and assume perfect recall. The detector catches the one event. Apply the one percent false-positive rate to the nine thousand nine hundred and ninety-nine non-events. The expected number of false alerts is ninety-nine point nine nine, about one hundred. Together, the true and false alerts give an expected total of one hundred point nine nine. Divide one true hit by that total. Precision is approximately zero point nine nine zero two percent, just under one percent. The marks round the expected queue to one true alert and one hundred false ones. An individual queue contains whole alerts and will vary; these decimals describe the expectation under our assumptions. Perfect recall has not made the checking work disappear.

04 · THE PERSONALITY SKETCH

Human judgment can underweight the prior as well. In their nineteen seventy-three study, Kahneman and Tversky gave two groups the same five personality descriptions. One group was told that thirty of the hundred professionals were engineers; the other was told seventy. Average estimates that a described person was an engineer were fifty percent in the first group and fifty-five percent in the second. The prior had a small, statistically significant effect. When no description was supplied, judgments matched the thirty and seventy percent priors. The study therefore distinguishes what people did with a description from what they did without one.

05 · INSENSITIVITY TO PRIOR PROBABILITY

In their nineteen seventy-four Science paper, Tversky and Kahneman used the heading, insensitivity to prior probability of outcomes. This is a claim about judgment. The rare-event calculation would still hold if every person understood it perfectly. An AI flag can arrive with a persuasive explanation, but persuasiveness supplies neither the event prevalence nor the detector's error rates. The question is whether the explanation carries evidence that changes the odds for this case.

06 · THE INTRUSION DETECTOR

Stefan Axelsson applied the calculation to intrusion detection in two thousand. His paper uses a hypothetical installation and explicit assumptions about its audit records and intrusions. Under those assumptions, a very low false-positive rate is needed for a substantial share of alarms to indicate intrusion. Each unresolved alert still asks for investigation. One record reaches the review desk while the others wait. Diane Vaughan's account of NASA describes how repeated anomalies became accepted practice. Extending that idea to routine dismissal of alerts is an organizational analogy. It does not show that a particular queue must cause fatigue or that the next true alert will be ignored.

07 · Advertisement · Bubble AI App Builder

Some ideas do not need another document before they become testable. With Bubble AI, you describe the app you want, and Bubble creates a working starting point: interface, data, and logic you can inspect. From there, you refine visually, connect AI models and services, and turn the first version into something real enough to use.

08 · CHOOSING THE THRESHOLD

A threshold selects an operating point. Raise it, and fewer cases qualify as alerts; that can reduce false alarms while losing true hits. Lower it, and more cases qualify, including both events and non-events. At the deployed prevalence, precision follows from the two class-specific rates. The operator has to weigh the cost of investigating a false alert against the cost of missing an event, and keep the required workload within available capacity. Some detectors offer an operating point that meets those conditions. Others need a different measurement or a different process.

09 · WHAT SURVIVES THE FLOOD

Requiring two signals to agree keeps only their intersection. A true event seen by the first signal alone is excluded. A false alert shared by both can remain, alongside a shared true alert. For fixed original flags, the combined rule cannot increase either the false-positive rate or recall. The question is how much of each it removes. Measure the second signal among cases already flagged by the first, separately for events and non-events. Multiplying the two marginal error rates requires conditional independence within each class. Different sources or model names do not establish that condition. Retained alerts can then enter a defined probe. Its result stays pending until measured, and a failed or inconclusive probe must remain possible.

10 · THE FIRST ENEMY

Before deploying the watch, estimate how many of its alerts will be true in the population it will serve. Record the cases, the outcomes and the follow-up needed to learn those outcomes. Check missed events separately; reviewing only the alert queue cannot measure recall. If the evidence does not support an answer, route the case for review. At the edge of a model's capability, an abstention can be useful, but the words I don't know are easy to produce. The policy needs validation: useful cases may be withheld and errors may still pass. The owner decides which combination of investigation, misses and delay the watch can bear.

01 / 10 · THE WATCHMAN'S ALARM0:00 / 7:02