Safe Perimeter · release v0.3
The alarm goes off,
and the adults who hear it come running.
A store makes a noise when a shirt goes out the door. Safe Perimeter makes a noise when a person under court-ordered monitoring crosses into a place built for children. He gets a private warning first and a chance to turn around. If he keeps coming, the alarm sounds where it happened, every adult standing there hears it, and the officer who already has authority over him knows in the same second. Nobody is named. Nobody watches the child.
What this is
A child walks to school, to the park, to a friend's house, and moves through the day inside a boundary that is awake. A man under court-ordered monitoring comes too close to it. He gets a private warning and a chance to turn around. If he keeps coming, the alarm sounds at that gate, the adults nearby turn around, and the officer who already has authority over him is notified in the same second. No name is ever spoken. No one watches the child.
We already do this for merchandise. Walk a shirt through a door without paying and an alarm goes off, and nobody argues about whether that is worth the cost. A child is worth more than a shirt. The perimeter lives in the places children already are: schools, playgrounds, daycares, pools, libraries. It lives in a shoe, in a band, in a phone. And it is built so that the system can be certain a monitored person is standing at a school door while remaining incapable of telling anyone who that person is until a court says so.
What the sound is for
A smoke alarm means get out. This one means the opposite: go toward it. When the alarm sounds at a playground or a school gate, it is telling every parent, teacher, coach and lifeguard within earshot that a person who is not allowed to be that close to children is that close right now, and that the right response is to walk over, be visible, and put adults between him and the kids. That is the whole mechanism. It is not a notification somebody reads later on a phone. It is a room full of adults turning around at once.
The sound is aimed at the adults who are present, not at the child, and it is deliberately hard to trigger, because an alarm that goes off for nothing is an alarm people learn to ignore. What it takes to earn it is set out on the alert tiers page: two independent sensors agreeing, a sustained thirty seconds, a place whose whole purpose is children, and a person who signs their name to the decision. When it goes off, it means something.
Tracking is the easy half, and much of it is already law. The half nobody has built is everything around it: who is enrolled and by what process, what the sensing can physically resolve, what happens in the seconds after a line is crossed, who is ever allowed to learn a name, and how a person earns their way out. All of it is answered here, in full, with working models a skeptic can run in their own browser.
Half of it is already law
Continuous GPS monitoring of people convicted of sexual offenses against children, with geofenced exclusion zones, is already on the books in a number of states. That means the hardest political fight has been fought and twenty years of case law is standing behind it. What follows is the half nobody has built yet - and the six constraints that decide its shape.
Read these first
Six decisions that shape how this gets built
Three are physical and legal facts the build is designed around. Three are mistakes that ended earlier programs and are engineered out of this one. Every one of them is a reason this version works where those did not.
Decision 1 · Empirical
We state exactly what this covers, so nobody can use it against us
94.1% of child molestation arrests are of first-time sex offenders.[4] For victims under six, 48.6% of offenders against victims under 6 were family members, and 48.3% acquaintances - 3.1% strangers.[3]
A registry-triggered perimeter does one job: it keeps supervised people away from the places children gather, and it proves it. It is not sold as a shield against a relative in the house, because the honest number is public and any opponent would raise it in the first hearing. Saying it first is what makes the rest of the claim credible.
Decision 2 · Physical
Half the hardware already exists, and the shoe carries the other half
The supervised person is already fitted with an active, powered, GPS ankle unit. That side needs nothing new. On the child's side the ranging module goes in the sole of the shoe, because that is where a battery fits and nothing can be forgotten or pulled off, and the band carries the warning the child can feel.
Passive festival-style tags are ruled out on physics - 20 mm is the certified operating volume for passive NFC as of Certification Release 15 - the requirement was 15,240 mm.[13] against a fifty-foot requirement - and signal strength alone is no better: ±10 dB of movement in the reading comes from simply changing which person sits in the seat - against about 6 dB of real signal between 15 m and 30 m.[15] So the enforceable distance is settled by a mains-powered anchor at the places children actually are.
Decision 3 · Legal
Tokenization is a privacy control, not a legal authorization
Under Grady[6] the search occurs at the moment a device is attached and tracks. It is complete before any datum is stored, hashed, encrypted or discarded. How the data is handled afterwards does not decide whether a search happened - it goes to reasonableness, and it is one factor among several.
This is the most consequential correction in the whole document, because the original plan treated tokenization as the thing that made the tracking lawful.
Decision 4 · Operational
The alarm is engineered to be believed every single time it sounds
95.50% of monthly GPS monitoring alerts in Cook County were false, up from 57.79% under radio frequency.[17] 80% of alerts were never checked by officers in a Tennessee audit.[29] 1% of roughly 63,000 monthly California alerts were restricted-zone entries.[30]
An audit records an offender whose repeated tamper alerts each fell under the escalation threshold; the officer was never notified, the device came off, and a rape and murder followed.[31] A system that cries wolf is not merely annoying. It is the mechanism by which the real alert is missed. Every gate on the tier ladder exists so that when this alarm sounds, the adults who hear it move.
Decision 5 · Design
The alarm says come here, and it never says who
A loud public alert is instantly de-anonymizing: everyone looks up, and the person walking away is the answer. Cryptography does nothing about that.
So the first move is a warning delivered privately to the monitored person, giving him a chance to leave, and it is triggered by the zone rather than by a child, so it can never become a detector handed to exactly the wrong person. If he does not leave, the alarm sounds anyway. It tells the adults present to come toward it. It does not tell them a name.
The self-clearing warning, and why its wording never varies >
Decision 6 · Legal
Imposed at sentencing, which is the version that holds up in court
Courts have tolerated retroactive registration.[7] They keep drawing the line at retroactive monitoring and geographic restriction.[8] The version that survives is imposed prospectively at sentencing, with an individualized-hearing path for existing registrants rather than a blanket sweep.
Slower to reach existing registrants, and it is the difference between a program that runs and a program that gets enjoined in its first year.
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%.
Hwang, Maniar, Smith & Nivens (2026), "False Alerts Related to GPS Electronic Monitoring in Illinois Cook County Sheriff's Office," Criminal Justice Review; urban GNSS error, Satellite Navigation (2025)
Driven by GPS drift; the worst zip codes were distinguished by tall buildings.
What survives
Watch everything. Learn nothing you were not authorized to learn.
Set membership is local and real time. Identity is remote, authorized and logged. That single sentence is the architecture, and the console is built so the code enforces it rather than a policy promising it.
What is inside
- Coverage - what share of harm this can touch, with every assumption exposed as a control.
- Anchored zones and sensing - the physical constraint, and the three modalities that survive it.
- Alert tiers - the response ladder as a running classifier.
- The warning - a compliance opportunity that must never become a detector.
- Data flow - the key hierarchy, derived live in your browser.
- Devices and the bracelet - monitoring hardware, and a voluntary attestation wearable that produces evidence for the person wearing it.
- The console, isolation and identity - a working operations console, the trust domains it runs in, and the two-party gate on learning a name.
- Legal architecture and due process - the constitutional map, and the parts every failed program skipped.
- What it solves - the honest case, and a table of the claims this can and cannot support.
- Unsolved - nineteen problems, sorted by whether they block the project, limit it, or cannot be fixed at all.
- What to add - the roadmap in priority order, with a model for the largest unknown in the concept.
- Anchor security and accessibility - two named holes in the specification, closed.
- Evaluation - how this gets measured, and what would count as it failing.
- The full specification - 146 numbered requirements in one printable document.
- Open questions - what this does not resolve, and what must be checked before anything is filed or funded.
- Check - every instrument here, run against its vectors in your own browser.
On the instruments
Most of these pages carry a working model rather than a picture of one. They compute in your browser, make no network requests and store nothing. Where a model rests on an assumption, the assumption is a control you can move. The check page runs them all against their vector tables so you can watch them agree with what the pages claim.
Why this one can actually pass
Work in child protection normally forces a choice between safety and civil liberties, and then loses whichever constituency it did not choose. This does not make that choice, because it does not have to. This design is more protective of the monitored person's privacy than the system it would replace. Current GPS monitoring builds a continuous, queryable, breachable movement history held by an agency or a vendor and browsable by insiders. This holds none of it: routine operation produces no location history at all.
That is a structural property rather than a talking point, which means the same proposal can be put to a child-safety committee and to a civil-liberties organization without changing the pitch. Very little in this category can do that, and it is the single biggest reason a bill built on it might actually pass. The argument is made in full at what it solves.
Ready before the first dollar
Every hostile reader this will ever meet - a defense lawyer, a civil-liberties litigator, a skeptical sheriff, a procurement officer - has already been let loose on it here. Nothing left standing is guesswork. That is what makes it buildable.
How this works
1:11 · 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 is | This is a design document for a child proximity protection system. Not a product and not a pitch. It is written to survive a hostile reading, so it leads with what breaks. |
| 0:13 | Six findings | Six findings reshape the concept. Three are constraints the design has to be built inside. Three are failure modes that have already ended real programs. |
| 0:23 | The constraints | The people a registry can watch are not the people who do the harm. The wristband as described cannot be built. Tokenization is a privacy control, not a legal authorization. |
| 0:35 | The failure modes | Alert fatigue is the failure mode, not a usability problem. A public alarm points at the person the tokens protected. Retroactive imposition is where this fails. |
| 0:46 | What is left to build | The tracking half is largely built and already law. What remains is enrollment, sensing, the seconds after a threshold is crossed, who may learn a name, and how someone gets out. |
| 0:58 | What to do with it | None of the six is a reason not to build. All of them change what gets built. Most pages here carry a working model rather than a picture of one. |