Report 179 · Luxury Home Security
Estate perimeter systems (fence sensors, beams, AI camera analytics) are usually sold on two numbers: how reliably they detect a person and how rarely they cry wolf. Both numbers can be true and still mislead. A federal specification written for the hardest perimeter there is, a prison fence line, shows what those numbers mean when a buyer insists on defining them, and what a homeowner should ask before trusting one.
There is no consumer standard for perimeter detection on a private estate. There are, however, buyers who have been specifying perimeter sensors for decades because a missed detection is not an option for them. One of the clearest public examples is Correctional Service Canada's engineering specification ES/SPEC-0400, "Perimeter Intrusion Detection System for Use in Federal Correctional Institutions," Revision 2, dated 4 February 2002. It is a procurement document, not a residential guide, and it is over twenty years old. But it is unusually explicit about the vocabulary, and the vocabulary is where residential claims get slippery.
The spec sorts every alarm into four bins: genuine escape attempts, alarms caused by inmate or guard activity, false alarms, and nuisance alarms. The last two are the ones that matter here, and the spec defines them precisely.
A nuisance alarm is one "assessed as caused by the atmospheric environment, birds, small animals, ground and air vibrations and other assessable causes." In other words, the sensor responded to something real that wasn't a person.
A false alarm is different. If an alarm "cannot be attributed to any known cause by the operational staff or after detailed analysis of alarm records, it shall be classified as a False Alarm." And the spec says alarm causes are assessed by the institution's own operational staff, only.
That distinction is the first thing to listen for in a sales pitch. "Almost no false alarms" can be literally true of a system that wakes you up every windy night, because by these definitions wind is a nuisance, not a false alarm. And the classification depends on someone actually looking at each event. A system with no one assessing alarms can't honestly report either number.
This is the clause that changed how I read every alarm-rate figure. Under the spec's conventions for calculating alarm rates, when one sensor alarms "at a rate of 10 or more per hour," that period "shall be considered as downtime for the sensor and the alarms shall not contribute to the alarm rate calculation."
Read it twice. In the worst periods, the alarms are removed from the alarm rate and moved to a different ledger, availability. The spec also lets operators mask a sector "during periods of high alarm rates," and in its availability section it counts both a 10-per-hour rate and an applied mask as a sensor failure.
Inside that spec, this is honest bookkeeping. The contractor still has to hit both numbers: a weighted availability of no less than 99% over the warranty period, and alarm-rate ceilings of, on average, no more than eight nuisance alarms per 24 hours for the complete perimeter and no more than one false alarm per 24 hours per sensor. The problem comes when someone quotes one ledger without the other. A vendor's alarm rate can look excellent precisely because the bad hours were counted as downtime, or because somebody masked the zone that kept going off. The number that tells you whether the sensor was actually on is availability, and it's rarely in the brochure.
The spec does not accept a detection probability from a data sheet. The required probability of detection must be met "under all environmental conditions," and it must be "demonstrated prior to acceptance of the installed equipment" through an approved test program at the site. The fence sensor gets at least three randomized climbs per sector plus extra climbs at posts, corners and sector overlaps, and simulated cutting. The ground sensor gets at least five crossings per sector. Any spot where detection falls short is "treated as a dead zone" and fixed at the contractor's expense.
The acceptance math is stated too: "A minimum of 20 tests with no failures is required to permit acceptance," and three or more failures within 50 tests means the specified performance hasn't been achieved. For a sense of scale (my arithmetic, using the method I walked through for bomb robots in Report 172): a sensor whose true detection rate is 90% would pass twenty straight tests only about 12% of the time. Twenty clean runs is a real bar. A two-minute demo walk in front of a camera is not.
The spec also makes the contractor declare, in writing, "any known limitations of sensor detection capability," including intrusion methods, maintenance such as snow removal, and environmental effects such as frozen ground. Separately, for alarm rates, the bidder must give documented evidence from other installed sites and contact details for those users. That is the right instinct: the best evidence for how a sensor behaves on your property is how it behaved on someone else's, verified by them.
The spec requires two detection layers, a fence-disturbance sensor and a ground-movement sensor, and requires them to be complementary. The conditions that cause nuisance alarms in one "shall not be likely to cause nuisance alarms in the other," the two must respond to different physical stimuli, and "different techniques and equipment are required to successfully evade detection by each sensor."
That is the single most useful design idea in the document for an estate. Two cameras running the same analytics are not two layers. They share the same weather problems, the same lighting problems and the same blind spots. A camera plus a buried or fence-mounted sensor fail for different reasons, which is the point.
Two more clauses are worth borrowing. A power failure, processor fault, or cut or shorted cable "shall result in an output at the equipment interface that represents a continuous detection or alarm condition." Silence is never allowed to mean "all clear." And where a sensor can be jammed into alarming or into going blind "without observable intrusion," the spec requires that "such jamming shall be detected" and reported. As an electronic warfare officer, I'd put that clause on every wireless perimeter product sold today.
I help design the AI security systems for a veteran-owned (SDVOSB) home-security company run by fellow veterans. I do not own that company and earn nothing from this link. Full policy here.
AI camera analytics are mostly a nuisance-alarm technology. Their pitch is telling a person from a deer, a branch or a headlight sweep. That is an old problem. A 1988 Sandia National Laboratories paper on exterior video motion detection noted that older systems "tend to have high nuisance alarm rates" and that this "made them generally unacceptable for use as an exterior sensor." The technology is very different nearly four decades later. The question a buyer should ask hasn't changed: show me the nuisance rate, the false rate and the availability, measured on a site like mine, with someone actually assessing each alarm.
The probability-of-detection numbers live in companion standards I did not read. ES/SPEC-0400 points to ES/STD-0401 and ES/STD-0402 for the required Pd and confidence. I have not opened them, so I quote no Pd target from the spec.
This is a 2002 revision of a correctional spec. It is used here as a model of how a demanding buyer defines terms, not as a current requirement for anyone, and later revisions may differ. Residential and estate systems are not required to meet anything like it.
The Sandia source is an abstract. I read the published abstract of the 1988 paper, not the full test data, and I make no claim about any modern analytics product.
Perimeter detection is sold on detection and false-alarm numbers, but those numbers only mean something with definitions attached. A demanding spec separates false alarms from nuisance alarms, moves runaway alarm periods and masked zones into availability, demands on-site proof of detection, and requires sensors that fail differently and fail loudly. Ask for all of it. A system that can only show you one flattering number is showing you the ledger it chose.
Scope note: this report explains how a published perimeter-detection specification defines alarm and detection performance and what a buyer can ask. It contains no instructions for defeating any sensor, evaluates no specific product, and is not a substitute for a qualified security designer assessing a specific property. Disclosure: the author helps design AI security systems for a veteran-owned home-security company, as stated in the body of this report, and does not own that company.
U.S. Army combat veteran (Counter-IED / Electronic Warfare), peer-reviewed researcher in microwave spectroscopy, and founder & CEO of Shroombiosis. Consults on laboratory operations, AI, and supplement formulation.