The doorman who only writes when someone walks in.
Think of a building doorman with a logbook. He doesn't write "nobody came by" every minute: he only writes when someone enters or leaves. On a quiet night the page stays blank. Whoever opens the book in the morning and sees a full page at 3 a.m. knows, without having seen anything, that something happened at three.
Our telemetry compressor, TUBE, works the same way. It only stores a new value when the previous one no longer serves, that is, when the measurement drifts from it by more than the agreed tolerance. While the signal stays inside that band, it stays silent, and that silence is what makes the file small.
The same silence becomes a detector. When the signal keeps leaving the band, the compressor starts writing non-stop. "How much it had to write" is, in practice, a measure of surprise.
Three steps, and no trained detector.
1. The band. The compressor guarantees that no reconstructed measurement drifts from the original by more than the tolerance. To keep that promise, it draws a band of that width around the last value it wrote.
2. The silence. Real signals spend most of their time quiet, inside the band. That silence is where the space savings come from: bytes are only spent when the signal actually moves more than agreed.
3. The surprise. When an event arrives, the signal breaks the band again and again and the write rate shoots up. Counting writes per time window is already the detector. It costs nothing extra, because the compressor was running anyway.
That is why guaranteed compression and anomaly detection are, deep down, the same thing seen from two sides.
The band was already written down. So the alarm is proof, not a hunch.
An ordinary detector says "the model found this odd". Here the sentence is different: "the signal provably left the band of normal behavior". The difference is that we did not invent the band.
In each field it already exists in a document. The synchrophasor standard, for meters that track the power grid in real time, allows at most 1% error in the measured vector (IEEE C37.118). Grid frequency has a band set by the grid code. A freight contract rounds prices to half a cent. A robot has a position tolerance in its manual.
Because the compressor guarantees that no measurement goes beyond that band, when it has to speak it is because the signal left it. Detection inherits the guarantee of compression.
The same principle, in four unrelated worlds.
An internet cable on the sea floor. Light traveling through the Curie cable, between California and Chile, changes polarization when the ground shakes. On the day of the magnitude 7.4 earthquake in Oaxaca, Mexico, the compressor went from silent 98% of the time to 4%, and the write rate jumped from about 3% to 100% at 15:37 UTC, roughly eight minutes after the quake, when the waves arrived. Read the case.
The power grid. The frequency of the British grid, measured every second for six months, stays in a narrow band: the compressor was silent 94.7% of the semester. Each generator trip breaks the band, and the write rate rises 14 times above rest. Read the case.
An optical fiber that becomes 101 sensors. With a technique called DAS, each stretch of a fiber works as a vibration sensor. In a public recording of a seismic event, silence fell from 81% to 31%, and the map of writes lit up following the wave front along the cable. See the demo.
AI reasoning loops. Models sometimes get stuck repeating the same thought. SIEVE's complete search catches all three kinds of loop (literal, cyclic and paraphrased) without letting any through. Read the study.
| Field | Data | The band comes from | What happened | In practice |
|---|---|---|---|---|
| Geophysics | light polarization on the Curie cable (Zhan, 2021) | the cable's normal behavior | silent 98% → 4% | the earthquake shows up as silence collapsing, at the right time |
| Power grid | British grid frequency, 1 s, 6 months | the grid-code band | silent 94.7%; 14× in the event | each generator trip becomes a clean spike of writes |
| Optical fiber (DAS) | 101 sensors along the cable (DASCore) | the fiber's quiet behavior | silent 81% → 31% | the event shows up where and when it passes |
| AI | reasoning of models that get stuck | repetition of its own text | 3 kinds, none escapes | literal, cyclic and paraphrased loops are caught |
A whale call goes unnoticed.
The free detector works for strong, broad events: earthquakes, grid failures, seismic waves. They shake the whole signal and break the band unambiguously.
A fin whale call on a fiber cable, however, is a weak, narrow signal: a train of pulses around 20 Hz in the middle of a lot of noise. The band breaks everywhere because of the noise, and raw surprise cannot isolate the whale. We measured it: the correlation between surprise and the calls was 0.11, almost none. The same goes for partial discharge on a power cable.
For these cases you first have to filter the frequency band the signal lives in, and then detection works again, with proof. See the whale case. Saying where the method fails is part of the method. Try option (b) in the specimen at the top.
When to use the free detector.
Have the tolerance written down
A standard, a contract or a manual. Without it, the alarm goes back to being an opinion.
There must be silence to break
A signal that is quiet most of the time. If it vibrates all the time, there is no rest to compare against.
Strong, broad event: use it raw
Earthquake, grid failure, seismic wave. Counting writes is enough.
Weak, narrow event: filter first
Select the signal's frequency band and only then count. That's what worked with the whales.
How much each measurement may vary, chosen by whoever uses the data. Here it is also the width of the band.
The share of time the compressor has nothing to write, because the signal stayed inside the band.
How much the compressor had to write in a time window. When it spikes, something changed.
Where this could be wrong.
Four domains are not all of them
We measured on a subsea cable, the power grid, DAS fiber and AI reasoning. Other signals may not have enough silence for the method to work.
The alarm threshold is a choice
How many times above rest counts as an event is a decision. Move the tolerance in the specimen to see false alarms and missed events trade places.
Weak signals need a filter
Raw surprise failed with the whale (correlation 0.11). Without band selection, weak events slip through.
Detecting is not explaining
The alarm says the signal left the band, with proof. Why it left is another question, for whoever knows the equipment.
← Certified telemetry · stickybit.com.br
- Curie cable: Zhan et al. (2021), light polarization on a subsea cable; M7.4 Oaxaca earthquake. Demo: curie-eq.
- Power grid: British system frequency (National Grid ESO), 1 s, Jan–Jun 2026. Demo: gbfreq.
- DAS fiber: DASCore public registry (
dispersion_event.h5). Demo: das-eq. Fin whale: OOI North Cable. Demo: whale. - Reasoning loops: Doom loops: why only a complete search catches all three.
- Synchrophasor standard: IEEE C37.118 (total vector error ≤ 1%).