The honesty problem
Most of it happens in rooms like this one
Not on a street, not near a school gate, and not by anyone a monitoring program has ever heard of. A proximity alert cannot reach the majority of child sexual abuse, and a design document that will not say so out loud gets dismantled by the first informed reader.
This page exists to stop the concept over-claiming. The number it produces is small. That is an argument for scoping the claim precisely - not for abandoning the work.
Who actually harms children
Two figures set the ceiling on everything else in this design document, and the ceiling is low.
- 94.1% of child molestation arrests are of first-time sex offenders.DOJ SMART Office, SOMAPI ch. 8, restating Sandler, Freeman & Socia (2008), Psychology, Public Policy, and Law 14(4)
The DOJ SMART Office restatement of Sandler, Freeman & Socia. A Florida replication found 93.5%. [4] - 3.3% of released child molesters were rearrested for a new sex offense against a child within three years.BJS, Recidivism of Sex Offenders Released from Prison in 1994, NCJ 198281 (Nov. 2003) [2]
Sandler, Freeman and Socia analyzed two decades of New York State arrest data and found that over nineteen in twenty of all sexual offense arrests involved first-time sex offenders. A Florida replication by Levenson and Zgoba reached the same place from the other direction. The Department of Justice's SMART Office restates the child-specific figure, and a Florida replication reached the same place independently.
Hard constraint · coverage
A registry-triggered monitoring system is structurally incapable of intercepting roughly nineteen of every twenty sexual offense arrests, because the person is not on the registry when they offend. It also cannot fire when the offender is a household member, which is the most common case for the youngest children.
Who the offender is to the child
The relationship figures matter as much as the conviction history, and they move sharply with the age of the victim [3]:
- 48.6% of offenders against victims under 6 were family members, and 48.3% acquaintances - 3.1% strangers.Snyder, BJS, Sexual Assault of Young Children as Reported to Law Enforcement, NCJ 182990 (2000). NIBRS 1991-96, 12 states
- 4.7% of offenders against victims aged 6-11 were strangers; for ages 12-17 it is 9.8%.Snyder, BJS, Sexual Assault of Young Children as Reported to Law Enforcement, NCJ 182990 (2000). NIBRS 1991-96, 12 states
What changed on this page
An earlier version of this page used a single set of aggregate shares for all victims under eighteen, and a first-time-offender figure drawn from all sexual offense arrests rather than from child molestation arrests specifically. Both were defensible and both were less precise than what is above. The age-banded figures matter because they change the argument: for the youngest children, a family member and an acquaintance are each about as likely as the other, and a stranger is a rounding error. For teenagers the stranger share is roughly three times higher - still small, but no longer negligible. A system aimed at one of those populations is not aimed at the other.
The conclusion did not change, and it is not being softened. The better figures make the ceiling lower, not higher.
The superseded figures are kept here rather than deleted, so that anyone who read the earlier version can see exactly what moved and by how much:
- 34.2% of victims under 18 in sexual assaults reported to law enforcement were assaulted by a family member.Snyder, Sexual Assault of Young Children as Reported to Law Enforcement, Bureau of Justice Statistics NCJ 182990 (2000)
- 58.7% were assaulted by an acquaintance.Snyder, BJS NCJ 182990 (2000)
- 7.0% were assaulted by a stranger.Snyder, BJS NCJ 182990 (2000)
This is the share of cases REPORTED TO LAW ENFORCEMENT, which is not the same as the share of cases that occur. - 95.9% of sexual offense arrests over a 21-year period in New York were by people with no prior sex offense conviction.Sandler, Freeman & Socia, Does a Watched Pot Boil?, Psychology, Public Policy, and Law 14(4):284-302 (2008)
One state, one period. It is the most-cited figure of its kind, not a national constant.
Note what the shift does to the argument. On the all-ages figures a stranger accounted for one case in fourteen; on the under-six figures it is one in thirty-two. The category a proximity system is best placed to reach got smaller, not larger, when the evidence got better. That is the direction corrections usually run in this area, and it is the reason the claim on this page is scoped as narrowly as it is.
Why family contact scores zero, not "less"
A proximity rule works by noticing that two people are near each other when they have no reason to be. A parent, step-parent, sibling or household member has every reason to be near the child continuously and lawfully. There is no threshold that separates abuse from ordinary family life, so the model scores this category at zero rather than at some small number. If it scored it at anything above zero it would be inventing capability.
Coverage model
Move any assumption. The arithmetic is the same arithmetic used above.-
of child sexual abuse cases reported to law enforcement could, in principle, produce an alert from this system.
The model multiplies four independent fractions and does nothing else. Its defaults are now the child-specific figures rather than the all-ages ones. Two weaknesses are worth stating: the first-time-offender rate is measured across arrests, and the stranger-abduction subset plausibly has a higher prior-conviction rate than the general population of offenders, so the stranger figure here may be an underestimate; and under the anchored-zone architecture the third control is no longer "children carrying a device" but the share of at-risk contact that happens inside a place with an anchor in it, which is a deployment question nobody has modeled yet. The model cannot correct for either without a source, so it does not pretend to.
The right way to say it
A small share of cases, and the worst ones
Low frequency is not low value. The slice a proximity alert can reach is the slice with the worst outcomes.
About 115 stereotypical kidnappings - taken by a stranger or slight acquaintance and held overnight, transported a distance, ransomed, killed, or intended to be kept - were estimated in a year.Finkelhor, Hammer & Sedlak, NISMART-2, Office of Juvenile Justice and Delinquency Prevention (2002)
A 1999 estimate with wide confidence bounds. Later NISMART work put the figure in the same order of magnitude.
That is a very small number against the scale of child sexual abuse overall, and it is the category with the highest rate of serious injury and death. Systems aimed at rare, catastrophic events are ordinary in every other field of safety engineering. Nobody argues that because structural fires are rare, sprinklers are a poor investment.
So the claim to make is this one, and only this one:
The defensible claim
Safe Perimeter closes a specific gap: it detects corroborated, sustained presence of a person under active monitoring inside a classified protected place, and it acts at the moment of encounter rather than at the next case review. It does not detect abuse by family members, by trusted acquaintances, or by anyone who has never been convicted - which together are the large majority of cases.
Stated that way, the concept survives a hostile reading. Stated as a solution to child sexual abuse, it does not survive the first paragraph. This closes a specific gap. It is not a solution to child sexual abuse, and no page on this site will say that it is.
The harm the over-claim causes
A device that stays silent while an offense is occurring - because the offender was never on the registry - can cause a parent to relax supervision that was doing more work than the device ever will. That is a second-order harm produced by the marketing rather than by the engineering, and it is the objection child-safety professionals raise first. Product messaging has to actively counteract it, not merely avoid making it worse.
One more thing the physical layer does not touch
Grooming of older children has largely moved online. It does not require physical proximity to begin, and by the time it does require proximity the relationship is usually already established and the meeting is consensual from the child's point of view. A proximity alert has nothing to say about any of it.
1.4 million online enticement reports were recorded by NCMEC in 2025.NCMEC CyberTipline data (2025)
NCMEC attributes much of the rise to expanded mandatory reporting under the 2024 REPORT Act. This is report volume, not measured incidence. [37]
That figure has to be handled carefully, and the caveat above is not a formality. It counts reports, not incidents. Read as a measure of how much online enticement is happening, it would be wrong in a way that flatters whoever quotes it. Read as what it is - the volume arriving at one reporting body under a widening reporting duty - it establishes only that the digital channel is where a great deal of the attention now goes, which is enough to make the point: this design document does not touch it, and should not be sold as though it did.
This is worth naming explicitly rather than leaving as an unstated gap, because it is the most common form the harm now takes among teenagers, and because 2017 the Supreme Court struck down a ban on registrants accessing social media. - the online layer has its own constitutional constraints, and they are not the same ones.
Packingham v. North Carolina, 582 U.S. 98 (2017)
How this works
1:09 · narrated · captioned · nothing here is sent anywhere when you press play. Prefer to read it? The transcript is below.
Transcript
| At | Section | What it says |
|---|---|---|
| 0:02 | What this page is | This page exists to stop the concept over-claiming. The number it produces is small. That is an argument for scoping the claim precisely, not for abandoning the work. |
| 0:12 | The ceiling | Ninety four percent of child molestation arrests are of first-time sex offenders. They are on no registry when they offend, so a registry-triggered system cannot see them at all. |
| 0:23 | Zero by construction | For victims under six, about half of offenders are family members. Family contact scores zero here, not less, because no threshold separates abuse from ordinary family life. |
| 0:34 | The model | The model multiplies four fractions and does nothing else. Every assumption is a control you can move. Move them honestly and the ceiling stays low. |
| 0:44 | The defensible claim | So the claim is narrow. It detects corroborated, sustained presence of a monitored person inside a protected place, and acts at the encounter rather than the next review. |
| 0:55 | The over-claim harm | Low frequency is not low value. But a device that stays silent can lead a parent to relax supervision that was doing more protective work than the device ever will. |