Four words for this page.
What the system says about itself: the position in the report, the measured number, the "I'm working". Whoever generates it picks every piece, and changing it costs nothing.
What comes out alongside without anyone choosing it: the pitch of a radio, the fuzz of a sensor, the size of a file, the time each packet arrived.
The minimum a live sensor cannot help emitting. Below it, what arrives is not measurement; it is repetition or fabrication.
Refuted, no alarm and undecidable. "No alarm" is not a clean bill of health, and "undecidable" is an answer, not a failure.
What the person says, and the sweat.
In an interrogation there are two sources of information about the person talking. One is what they say: they pick every word, and lying is free. The other is what they cannot control while saying it: the sweat, the tremor in the voice. Almost every systems audit reads only the first.
A satellite reports its own orbit in the telemetry it sends down. Editing that number costs one line of code. But the radio signal arrives with its pitch shifted, like an ambulance siren that sounds higher coming and lower going (the Doppler effect). That shift is set by the real speed, not by the onboard software.
To lie about the pitch, as heard by several antennas in different places at the same time, you have to be in that orbit. Faking it costs the same as complying. Security people call this a side channel and usually use it to attack. Here, it is used to check.
Declarations became free. Physics did not.
A report, an assessment, a well-formatted number: with AI, producing any of them costs almost nothing, and costs the same whether it is true or not. Everything that is a declaration lost its price. Checking, on the other hand, still costs a person reading carefully.
The physical leak was not part of that clearance sale. A language model does not shift the pitch of a radio or make a still sensor tremble. To fabricate a machine's leak, you need the machine doing the thing.
This also explains why the same idea does not rescue text. A paragraph leaks nothing its author does not control: every letter is a declaration. Watermarking generated text is an attempt to create a leak by decree, and a leak by decree is a declaration under another name. The exception that proves the rule is a real camera photo: every camera sensor has tiny factory imperfections that are printed into every photo it takes, and a generated image does not have them.
Absence convicts. Presence does not acquit.
There are two kinds of leak, and they are not worth the same. The physical kind, like the satellite's pitch heard by several antennas, can only be imitated by doing the thing. That one can serve as a certificate: if the pitch matches at all five antennas, the satellite is where it says.
The statistical kind, like a sensor's fuzz, can be imitated with a random number generator, as long as the fraudster knows it is being measured. That one can only convict. If the fuzz is gone, something is almost certainly wrong. If the fuzz is there, it proves nothing: it may be the sensor, or someone who read the contract. That is what option c in the specimen shows.
In practice, the most common mistake is not a missing detector. It is reading a detector's silence as a certificate. A green dashboard says "nothing fired", not "everything is fine".
The dead sensor looked more normal than normal.
In August we tested one of our anomaly detectors against plausible faults, on public data from a robot arm: 55,000 readings from 14 joints. One of the simulated faults was the most treacherous: the sensor freezes on a normal value and keeps repeating it, with no error.
The detector scores how strange each stretch is. The result came out backwards: in every comparison, the frozen sensor was rated more normal than the live one. It makes sense once you see it. A live sensor, even with the machine quiet, still trembles a little; the frozen one emits exactly the same number, and nothing is more predictable than that.
The winner was a three-line rule: "the fuzz dropped below the minimum for too long". It does not look at the content; it looks at what the sensor cannot help emitting. The sophisticated detector failed because it asked "does this look normal?". The simple rule got it right because it asked "is this still leaking?".
Our bet lost, and the coarse sensor leaks nothing.
After that finding, we bet on a leak our telemetry compressor already produces for free: the size of each packet. A live sensor does not compress below the floor; a frozen one compresses to almost nothing. We wrote down the bet and the criterion for losing it before measuring, and ran it on the same robot arm.
The bet lost. Packet size was no better than the simple fuzz rule in any situation. And in the hard case, a sensor that freezes precisely while the machine is quiet, the best method got only two pairs in three right, and it was not our bet.
| Situation of the sensor that froze | Strangeness detector, read backwards | Minimum fuzz rule | Packet size (the bet) |
|---|---|---|---|
| The joint usually moves | 100% | 100% | 100% |
| The joint is usually quiet | 65% | 50% | 56% |
| Quiet, and its neighbours still too | 68% | 59% | 62% |
How to read it: in what share of pairs (one frozen stretch, one live stretch) each method correctly points to the frozen one. 50% is a coin toss. Our measurement, August 2026, with freezes and faint fuzz simulated on top of the real data. When a quiet joint freezes, "the sensor died" and "the machine stopped" produce the same data: the honest verdict there is undecidable, and that happened in 4.5% of the freezes.
And a physical limit showed up that applies to any product along these lines. The joints' position sensors measure in coarse steps. When still, they return exactly the same number, with no fuzz at all, because the tremor is smaller than one step. There is no leak to check, and a live, still sensor looks identical to a frozen one. That is option d in the specimen.
On the same robot's gyroscope, which measures in steps finer than its own tremor, the math works. The rule raised no false alarm in 34 real stretches, caught the 5 freezes we injected, and declared the coarse sensor undecidable instead of guessing. Saying where the instrument is blind is half the product.
A side finding, logged as an observation and not a result: on the coarse sensor, the leak changes target. The real sensor at rest spends almost nothing; whoever invents fuzz to look alive ends up spending more than they should. There, packet size does not catch the freeze. It catches the forger.
An auditor who only sees sizes.
Packet size has a property the fuzz itself does not: it is visible to someone who cannot see the content. An insurer, a regulator or an outside auditor can receive only the sizes and the times, timestamped, and establish "this sensor has been dead for 40 packets" without ever reading a measured value.
That matters in pay-per-use contracts, which today are settled on the word of whoever gets paid. One possible clause: "packets below the declared floor for N in a row: undecidable, and the period is not billable".
The rule already exists as a tool, with the three verdicts and the declared limit. What does not exist yet is a buyer. Until someone says this is worth a contract, it is a measured idea, not a product.
Where else the leak checks the declaration.
| Where | What is declared | What leaks unintentionally | What it lets you check |
|---|---|---|---|
| Ships | Position and identity on the onboard radio | Each transmitter's unique quirks, the pitch, the arrival time at several antennas | Whether it is that ship, and whether it is where it says |
| Cloud | "Your data stays in Brazil" | Response time: light cannot travel faster than the distance allows | Refutes that the server is where they say |
| Recordings | "Recorded on such a day" | The hum of the power grid, which wobbles in a unique way at every moment | Date and region, against the grid's records |
| Healthcare billing | "We ran N scans" | The equipment's own power draw and cycles | Refutes a charge without opening medical records |
| Published research | Averages and tables | The arithmetic that real data always obeys | An average impossible for that number of people signals fabrication |
Well-known examples from other fields, to show the pattern. We measured none of them; the numbers and limits of each are in each field's own sources.
Four questions before buying monitoring.
What is declared, and what leaks?
If everything the verdict reads was chosen by the audited party, there is no audit. There is a declaration with a stamp.
How much does faking the leak cost?
Compare it with the cost of simply doing the thing. A physical cost survives whoever reads the contract; a statistical cost dies when the fraudster finds out what is measured. And they do.
Where does the signal vanish?
15-minute averages swallow the fuzz, batched packets swallow the timing, coarse sensors do not tremble. If the vendor cannot name its blind spot, "100% coverage" is the first thing to doubt.
What is the third verdict?
An honest monitor answers refuted, no alarm or undecidable, and says how often it lands on each. A dashboard with only green and red is rounding the doubt one way or the other.
Where this could be wrong.
Fuzz can be invented
Whoever knows the fuzz is measured can generate fake fuzz. That is why a statistical leak only convicts. A real certificate needs a physical leak, and those are rarer and more expensive to read.
No buyer yet
Checking without opening the data works on the bench. What is missing is someone (insurer, regulator, auditor) saying it is worth a contract clause. Without that, it is an interesting property, not a market.
One robot, few sensors
The measurements come from one public robot-arm dataset, with faults simulated on top of real data. Other machines have other sensors and other blind spots; the floor has to be declared sensor by sensor.
The channel that checks also watches
The same leak that proves a machine ran reveals when it ran, even in "anonymised" data. Whoever sells the check has to treat that as a declared privacy cost, not a freebie.
← Research notebook · Certified telemetry · stickybit.com.br
- Our measurements, August 2026: the detector test against plausible faults and the packet-size-as-detector test, with the losing criterion recorded before measuring. Data: the public LeRobot ALOHA dataset (Hugging Face), 55,000 readings from 14 joints, with simulated freezes and fuzz.
- Camera sensor factory imperfections as a fingerprint: Lukáš, Fridrich and Goljan, "Digital camera identification from sensor pattern noise" (IEEE TIFS, 2006).
- Power-grid hum to date recordings: Grigoras, "Digital audio recording analysis: the Electric Network Frequency criterion" (2005).
- Impossible averages in published research: Brown and Heathers, "The GRIM test" (Social Psychological and Personality Science, 2017).
- Location from response time: Gueye et al., "Constraint-based geolocation of Internet hosts" (IEEE/ACM ToN, 2006).
- The specimen is our own simulation: signal, freeze and fuzz generated in the browser with a fixed seed; packet size is estimated from the variety of steps between consecutive readings, plus a fixed 2-byte header.