The sensing decision
The band works, and the shoe is what makes it wearable
The supervised person already wears a powered, court-ordered transmitter - that half is deployed in every state that does this. The child side listens and never announces itself. Anchors at the places children actually are confirm the distance.
Three parts do the sensing, and only one of them has to be invented. A person under court-ordered supervision is already fitted with an active, powered, GPS ankle unit that transmits continuously - that is not a proposal, it is the equipment already bolted to the population this system is aimed at. On the child's side there is a module in the sole of the shoe, where a battery and an antenna have somewhere to live and nothing can be forgotten on a countertop or pulled off in a hallway, and a band on the wrist for the alert the child can feel. And at the places children actually are - schools, playgrounds, daycares, pools, libraries - a fixed anchor on mains power settles the distance question that no battery-powered thing should be asked to settle alone.
Every part of that is an active radio, and it is active on purpose. Passive tags - the cloth NFC bands handed out at hotels and festivals - are ruled out here, and it is worth showing why so nobody proposes them later as a cost saving. A passive tag backscatters a reader's field; it never originates one. The NFC Forum's Certification Release 15 expanded the mandated operating volume from 5 mm to 20 mm and that was reported as a significant advance [13]. The distance this system cares about is 15.24 m. Passive is not a near miss and no amount of antenna work closes it, which is the entire reason the child side is a powered module in a sole rather than a printed tag in a strap.
Why the sole is the engineering answer and not just a comfort feature
Active ranging at fifty feet needs a battery, and a battery is exactly what a wrist strap has no room for. Table 0.2 below has one technology that meets the requirement outright - ultra-wideband, at under 15 cm - and the only thing it asks in return is power and a two-way exchange. Put that module in a shoe sole and the objection evaporates: there is volume for a cell, the antenna sits away from the body, the device is worn for every hour the child is outside the house, and there is no clasp to defeat. The wrist keeps the job it is actually good at, which is telling a child something on their own skin without anyone else hearing it.
This is also what makes the manufacturing route real. A sole cavity is a tooling change to a product line that already exists, which is a conversation a footwear company can have. A new object a child has to remember is not.
What the anchors are for: signal strength alone cannot resolve these distances. Two peer-reviewed real-world studies of COVID exposure notification measured exactly this, at a two-meter threshold - a far easier problem than fifty feet. Deployed national configurations on a light-rail tram produced true-positive rates ranging from nothing at all to about half, the best of them paired with an equally large false-positive rate. On a commuter bus where every one of sixty handset pairs was inside two meters for more than fifteen minutes - so every pair was a true exposure by construction - detection ran from nothing to roughly a third depending on threshold. The authors concluded performance was "similar to that of triggering notifications by randomly selecting from the participants, regardless of proximity" [15].
Leith & Farrell, PLOS ONE (2020) e0239943 (tram) and (2021) e0250826 (bus); MIT Lincoln Laboratory
The German configuration reached 9%; the Italian about 50%, with a matching 50% false-positive rate.
Table 0.2 - candidate sensing technology against a 50 ft requirement
| Technology | Best documented accuracy | Meets the requirement |
|---|---|---|
| Passive NFC (ISO 14443) - hotel and festival bands | 20 mm certified [13] | No - 762× short |
| Passive HF vicinity (ISO 15693) | 1-1.5 m | No - 10× short |
| Passive UHF (EPC Gen2) | 1-5 m, reader-powered | No - not peer-to-peer |
| BLE RSSI ranging | 0-50% true-positive at a 2 m threshold [15] | No - equivalent to chance |
| GPS position differencing (urban) | 18.4 m RMS per position; ~26 m on the separation, derived [17] | No - error exceeds the threshold |
| BLE Channel Sounding (Bluetooth 6.0) | NLOS p95 4.89 m [16] | Marginal - and needs power |
| Ultra-wideband (UWB) | under 15 cm 2D [18] | Yes - active, powered, two-way |
One row in seven says yes, and it says yes on one condition: ultra-wideband is active, powered and two-way. That is the row this system is built on. The requirement is 15,240 mm; the certified passive figure is 20 mm, which is why nothing in the middle of the table is worth revisiting.
Design decision · D-0
Active radios only, and the power budget goes where there is room for it
Ultra-wideband meets the distance requirement and needs power to do it. Rather than treat that as a reason to abandon a worn device, the power budget is placed where a battery physically fits and stays with the child all day: the shoe sole. The wrist band is an alerting surface, not the ranging radio. The supervised person's side needs no new hardware at all - the court-ordered ankle unit is already active, already powered, already transmitting. And the heaviest, most expensive half of the work, continuous receiving, is handed to a mains-powered anchor that never has to be charged by anybody.
The governing arithmetic
It is worth stating plainly, because it is the reason no amount of engineering effort rescues the first approach. Free-space path loss between 15 m and 30 m is about 6 dB - that is the entire signal the measurement has to work with, the whole difference between standing at the far edge of the requirement and standing at half of it. ±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]
Nuisance variance from a single variable already exceeds the entire usable signal, and orientation, pockets, walls and crowds have not been counted yet. That is not a filtering problem, a calibration problem, or a machine-learning problem.
What this design document said before, and why the weaker number is the stronger argument
An earlier draft of this page put a second figure alongside that one: measured nuisance variance from antenna orientation alone of 19 dB. The revised source specification withdraws it, and so does this page. It is not a figure the underlying measurements assert, and a number that cannot be traced back to a study is worth less than no number at all - it is the sentence opposing counsel reads out.
Removing it does not weaken the finding; it sharpens it. The argument no longer needs a stack of noise terms added together. One variable - which person is sitting in the seat, holding the phone the same way in the same place - already moves the reading further than the whole distance being measured. Everything else is on top of that.
No filtering recovers information that was never in the measurement
A Kalman filter, a rolling median, a trained classifier, a longer averaging window - each of these can reduce the variance of an estimate that contains signal. None of them can manufacture a distance from a reading in which distance was never encoded. Anyone who proposes to fix RSSI ranging with better software is proposing to average a coin flip until it becomes a ruler.
Leith & Farrell, PLOS ONE (2020) e0239943 (tram) and (2021) e0250826 (bus); MIT Lincoln Laboratory
Before accounting for orientation, pockets, walls or crowds. Nuisance variance from a single variable already exceeds the entire usable signal.
Move the hard work off the body
The distribution of work is the design. A worn device should not be asked to compute a distance it can be wrong about, so fixed anchors do the confirming ranging and the worn devices stay simple.
This works because the places that matter do not move. Schools, playgrounds, daycares, pools and libraries have addresses. An anchor mounted at such a site can be powered from the mains, carry four antennas, run ultra-wideband or Channel Sounding continuously rather than duty-cycled down to nothing, and be professionally sited and calibrated once by a person standing in the room. Every constraint that squeezes a worn device is relieved by moving the expensive half off the body - the monitored unit is left broadcasting, which is the cheap half, and the child's module is left listening, which costs the child nothing at all.
Table 3.2 - why the anchor topology resolves each constraint
| Constraint | Peer-to-peer wearables | Anchored zone |
|---|---|---|
| Power for continuous scanning | Continuous scanning is the expensive half - receiving costs about five times as much as transmitting, per bit [18] | Mains-powered anchor. The wearable only broadcasts. |
| Antenna diversity | One antenna on a wristband | Four antenna paths - dropping to one raises NLOS p95 error by about half [16] |
| Calibration | Impossible - the environment changes with every step | Site-surveyed once, at installation, by someone who was there |
| Urban GNSS error | 18.4 m RMS per position; ~26 m on the separation estimate, derived from it [17] | Not used for the proximity decision at all |
| Zone semantics | Unknown - is this a park or a pharmacy? | Known, and classified per anchor before it is switched on |
The last row is the one that is easy to skim past and hardest to get any other way. No worn device knows whether the room it is standing in is a nursery or a pharmacy. An anchor was installed by somebody who knew, and that knowledge is what the tier ladder needs in order not to fire on a supermarket.
Design decision · D-1
The monitored unit transmits; the child side listens and never transmits; anchors range, classify and decide
The court-ordered ankle unit broadcasts a rotating identifier and does nothing else - no ranging, no map, no conclusion. The anchor runs the two-way exchange, knows what kind of place it is standing in, and makes the classification. The child side - sole module and band - is receive-only: it can hear a monitored identifier nearby and warn the child directly, and it never emits anything of its own. That asymmetry is deliberate and load-bearing. A child-side device that transmitted would be announcing a child's position to anyone with a receiver, which is a stalking instrument. Listening costs the child nothing and gives the warning even in a spot no anchor covers.
The topology
Left is a battery-powered wearable being asked to do everything on its own. Right is how it is actually built. Same wearables, different distribution of work: the expensive, error-prone half moves to a box on a wall that has mains power, four antennas and a professional install, so the thing a child wears can be simple, silent and cheap enough to put in every pair of shoes.
The trade, stated before anyone else states it
Anchored zones protect places, not people in arbitrary locations. A monitored person and a child who meet on a footpath two streets from the nearest anchor generate nothing. That is a real loss of coverage and it should be in the first paragraph of any pitch, not the appendix.
It is also, on the evidence, the correct narrowing rather than merely a tolerable one. The coverage finding establishes that this system can only ever address a specific minority of cases: court-supervised individuals with a prior child-victim conviction, encountered in public and semi-public space. It cannot reach the household member, and it cannot reach the person who has never been convicted of anything. The addressable case is public and semi-public space near protected places - so a topology that covers exactly that and admits it covers nothing else is aligned with what the system could ever have done.
Blanketing everywhere would be worse in both directions at once. It would mean a monitored person's proximity is computed everywhere they go, against everyone they pass, which is a far larger intrusion than the one the case law will bear - and it is the version a court is most likely to strike down. Coverage that holds up in a hearing is worth more than coverage that gets thrown out.
Zone classes, assigned at installation
An anchor is not neutral infrastructure. What it is allowed to conclude depends on what kind of place it is bolted to, and that has to be decided before it is switched on.
Children's presence is the purpose of the place
- School grounds, during school hours - the same fence at nine on a Sunday evening is not a Class A zone.
- Playgrounds and equipped play areas.
- Licensed daycare and nursery premises.
- Public swimming pools and splash parks.
Only a Class A anchor can carry an encounter to the highest tier automatically. A monitored person inside one, for a sustained period, in a place whose entire reason for existing is that children are in it, is a fact that needs no further interpretation.
Children are present incidentally
- Public libraries.
- Shopping centers and supermarkets.
- Transit stations and interchanges.
- Parks with no dedicated play equipment, high streets, car parks.
A Class B anchor can log, and it can carry a silent notification to guardians enrolled in that zone. It cannot reach the top tier automatically, ever. Getting there from a Class B zone requires something no location can supply by itself: a human adjudication that names the reason on the record.
This is the rule that stops a monitored person buying groceries from generating a dispatch
A supermarket contains children at nearly every hour it is open. So does a bus. So does a library. If proximity in those places could reach the top tier, the system would refer a monitored person to law enforcement for doing the shopping - repeatedly, lawfully, and in a way no supervising officer could act on. The field record is unambiguous about where that leads: an alert stream that is mostly false is an alert stream that stops being read, and both effects are documented below.
| Deployment | Measured | What it shows |
|---|---|---|
| Cook County, GPS zone alerting | 95.50% of monthly alerts false, up from 57.79% under radio frequency [17] | Drift alone can make the majority of alerts wrong. The worst zip codes were distinguished by tall buildings. |
| Tennessee, statewide audit | 80% of alerts never checked by an officer [29] | What a false-alert rate does to human response is stop it. |
| California, monthly alert volume | 1% of roughly 63,000 monthly alerts were restricted-zone entries [30] | Almost all of the volume is noise even in a mature program. |
Classified, never inferred
An anchor's class is recorded at installation, by a person, against an address, in a register that can be inspected and appealed. It is never derived at runtime from what the anchor observes.
- Inference is a surveillance capability in disguise. An anchor that decides "many small devices move through here at half past three, so this is probably a school" has, by construction, built a model of children's movements. That is a system nobody agreed to install, arrived at as a side effect of a convenience feature.
- Inference is attackable. Anything a system concludes from ambient observation can be manufactured by whoever controls the ambient observations. A classification that can drift upward on evidence can be pushed upward by an adversary, and the top tier is the one that sends officers.
- Inference has no audit surface. A written classification can be produced in a hearing, checked against the address, and challenged by a person whose liberty turns on it. A learned one can only be described.
- Class A carries a time window, not just a location. School grounds are Class A during school hours and Class B outside them, because the same square of tarmac means something different at four in the afternoon and at midnight. The window is part of the record.
- Reclassification is a change-controlled event. Raising an anchor from B to A is a decision with the same weight as installing it, and it is logged as one.
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 the rest of this document inherits
- No passive tags anywhere. Every radio in the built design is active and powered, and the child's power budget lives in the shoe.
- An enforceable distance is something an anchor produces, from a two-way exchange it initiated - see the sensing layer for which modalities are permitted to produce it and which are forbidden. The child-side warning does not need that precision, because it warns a child rather than dispatching an officer.
- The classifier's zone input is an anchor's recorded class, not a guess made from a position fix.
- Coverage claims are bounded by anchor coverage, and must be stated that way in every funding document.
How this works
1:08 · 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 | The correction | This page carries the correction that reshapes the rest of the design. The wristband as first described cannot be built, and the reason is physics rather than effort. |
| 0:12 | Two passive bands | A passive tag has no transmitter. It backscatters a reader's field. Two passive bands cannot illuminate each other, so the pair has no detection range of its own. |
| 0:22 | Signal strength | Signal strength cannot rescue it. In real studies at a two metre threshold, detection ran from nothing to about half. The authors called it equivalent to choosing at random. |
| 0:32 | Off the body | So the ranging moves off the body onto fixed anchors. The places that matter do not move. An anchor can be mains powered, multi-antenna, and calibrated once at installation. |
| 0:44 | Zone classes | Every anchor is classified before it is switched on. Class A is a place whose purpose is children. Class B, where children are incidental, can never reach the top tier by itself. |
| 0:54 | The trade | Anchored zones protect places, not people everywhere. An encounter two streets from the nearest anchor generates nothing. That belongs in the first paragraph of any pitch. |