Analyst guide · Anomalous Phenomena
Direct answer: Sensitivity and the fraction of alerts that are true answer different questions. To interpret an alert, combine the supplied target count, missed targets and false alerts instead of repeating an accuracy slogan.
A practical verification workflow
- Define a target and establish the ground-truth population for the exercise. The word target should not quietly change from a test pattern to a verified threat or extraordinary contact.
- Make a two-by-two table: target with alert, target without alert, non-target with alert and non-target without alert. Check that its counts sum to the full dataset.
- Calculate sensitivity with actual targets as the denominator and alert precision with all alerts as the denominator. Label each fraction explicitly.
- Limit the conclusion to the dataset and conditions supplied. A new deployment with a different prevalence cannot simply inherit the same precision.
A separate worked illustration
Consider a separate invented test with twenty true signals. Eighteen trigger alerts, while eighty-two non-signal windows also trigger alerts. Sensitivity and the share of alerts that are true now use very different denominators. An analyst should report both, rather than presenting the larger-looking metric as the answer to every question.
The common reasoning error
A correct percentage can still mislead if it answers the wrong question. Ask whether the reader wants detection among targets or reliability among alerts.
Keep an auditable analyst note
Record the question, source identifier, original observation, conversion or calculation, and the narrow conclusion it supports. Separate a confirmed fact from an inference and from information that is still unknown. Keep the original timestamp or unit alongside any normalized value so another reader can reproduce the reasoning.
Practise with the related case
Operation Alert Denominator: Recalculate a Rare-Signal Detector applies this skill in a fictional evidence packet with five graded objectives. This guide explains the method, not the hidden answer key. Case evidence and entities are invented; the separate illustration above is not a substitute for reading the dossier. Sign-in is required to use the case terminal.
Questions readers ask
Does high sensitivity guarantee that most alerts are true?
No. False alerts and the base rate also affect the alert population.
Can reposting a detector screenshot improve its ground truth?
No. Repetition changes distribution, not the underlying test labels.
Reporting context and further reading
NASA impact-structure observations · 2026-10-13. This primary-source reporting provides recent topic context only. It is not evidence that the invented case happened, and its publication date may differ from the incident dates described.
Explore the Case Vault · Browse Intel Drops