“Not detected” sounds like a claim about the world. Usually it is a claim about the meeting between a method and a sample: under these conditions, with this instrument, search, threshold and time window, the target did not produce a reportable signal. That can be useful evidence. It is not the same sentence as “the target was not present.”
The distinction reaches far beyond laboratories. A security camera did not record a person. A search found no document. A sensor raised no alarm. A diagnostic test returned negative. A monitoring dashboard showed no error. In each case, the negative result inherits the limits of what was observed and how.
A result belongs to a measurement system
A detector does not divide reality into present and absent with perfect authority. It converts some property of the world into a signal, processes that signal and applies a rule for reporting. Weak signals can be indistinguishable from background variation. Strong signals can be missed if the wrong place, time, format or feature is sampled.
The NIST/SEMATECH e-Handbook of Statistical Methods treats measurement as a process with variation, calibration and uncertainty. The lesson is broader than any one instrument: the output cannot be interpreted responsibly without the process that produced it.
“Not detected” therefore has at least four possible explanations. The target may be absent. It may be present below the reporting threshold. It may be present outside the sampled region or time. Or the method may have failed or been aimed at the wrong signature. The result alone does not choose among them.
Thresholds create legibility, not metaphysical borders
A threshold turns a continuous and noisy signal into an action. Above the line, report or alarm; below it, do not. This is often necessary. Without a rule, every fluctuation could become a false alarm and every analyst could produce a different answer.
But the reporting line is not a wall in nature. A concentration slightly below a laboratory's quantitation limit is not magically zero. A motion event just below a camera's configured sensitivity is not nonexistence. A document excluded by a search index has not been disproved by an empty results page.
The U.S. Environmental Protection Agency's detection and quantitation procedures distinguish questions about detecting a signal from questions about measuring it reliably. The technical details are method-specific, but the reasoning travels: detection, identification and useful quantification are different jobs.
The sample defines the world the result can describe
A perfect test of the wrong sample answers the wrong question. Water from one tap does not represent every point in a plumbing system. A log search covering one account does not clear every account. A survey of current users cannot reveal the experience of people who already left. A camera pointed at the driveway cannot establish what happened behind the house.
Sampling boundaries can be spatial, temporal or representational. The system may observe only during configured hours. It may store only events that cross a trigger. It may index text but not images, attachments or deleted revisions. A negative result should name these boundaries explicitly.
Timing matters because many signals change. Testing too early or too late can miss a temporary condition. A quiet week does not establish a permanently stable system. The relevant question is whether the observation window overlaps the process you are trying to detect.
The negative-result interpretation matrix
| Question | What to record | What it prevents |
|---|---|---|
| What was the target? | The exact substance, event, file, behavior or failure signature | Claiming absence of a broader category |
| Where and when was observation possible? | Sample source, coverage, retention and time window | Generalizing beyond the field of view |
| What could trigger a report? | Threshold, sensitivity setting, query or decision rule | Treating below-threshold as zero |
| What could block detection? | Noise, format, occlusion, timing, permissions or instrument failure | Assuming the method had no blind spots |
| What did controls show? | Calibration, known-positive check, logs or test query | Mistaking a broken detector for a clean result |
| What action depends on the result? | Reassure, monitor, repeat, escalate or use another method | Using one negative beyond its decision value |
Search results have detection limits too
Digital search feels exhaustive because the interface is fast. It is not exhaustive unless the collection, indexing and query actually cover the material. “No results” can mean the record is absent. It can also mean the file was never indexed, permissions hid it, optical character recognition failed, the terminology changed or retention deleted the relevant period.
For consequential searches, preserve the query, scope, date and account. Search synonyms and identifiers. Inspect likely source systems directly. If the claim is important enough to publish or act on, a second route with different blind spots is more valuable than repeating the same search ten times.
A negative can still be strong evidence
The point is not that negative results mean nothing. If a method is highly likely to detect a target at a relevant level, the sample covers the decision boundary, the detector is functioning and the target remains absent across suitable observations, the negative result can meaningfully reduce probability.
Strength comes from the design, not the grammar. A well-powered experiment that fails to observe a predicted effect can challenge a theory. A correctly configured monitor that passes a known-signal test and remains quiet during the relevant window can support continued operation. A complete repository search with preserved history can make a missing record genuinely informative.
Repeated negatives are most useful when they add independent coverage. Running the same narrow method again may reduce random error, but it does not cure a shared blind spot. A second method, new time window or different sample can contribute more than another copy of the first observation.
Use a four-line negative-result statement
- Target: state exactly what was sought.
- Coverage: state the sample, time and accessible region.
- Capability: state the relevant threshold and known blind spots.
- Decision: state what the result changes—and what would trigger another method.
A calibrated sentence might read: “No matching event appears in the retained logs for the named account between 2:00 and 4:00 p.m.; logs exclude local activity that never reached the service, so device records must be checked before concluding no event occurred.” The longer sentence carries less certainty theater and more operational value.
The honest conclusion is rarely “nothing.” It is a bounded update: the world is now less likely to contain the target in the region the method could see. Keep that boundary attached.
Sources and further reading
- NIST/SEMATECH e-Handbook of Statistical Methods, for measurement-process, variation and uncertainty concepts.
- U.S. EPA: Procedures for Detection and Quantitation, for method-detection and quantitation documents.
- NIST Uncertainty Machine, an official tool illustrating measurement uncertainty through an explicit model.
END OF TRANSMISSION 046
Keep the question. Test the model.
Choose the narrowest claim the evidence can carry, then leave room for revision.