A compressor with a guaranteed error bound draws a tolerance "tube" around the signal and stays silent while nothing changes. When the signal breaks out of the tube, the compressor has to speak — and that "speech" is, in practice, a measure of surprise. The event is detected for free, and certified, because the tolerance (the ε) already lives in a standard or a contract.
It is not "the model found it odd." It is "the signal provably left its normal band of behavior." We measured the same principle across four domains with nothing in common — and it worked in all four.
TUBE stores a new value only when the old one no longer serves — while the signal stays inside a tube of radius ε, it writes nothing. The reconstructed error is mathematically bounded by ε.
Real signals spend almost all their time quiet, inside the tube. That silence — the "deadband" — is what yields the compression: you only pay bytes when the signal actually moves more than the tolerance.
When an event arrives, the signal breaks the tube and the emission rate spikes. "How much the compressor has to speak" becomes an anomaly detector — free, because it was already compressing.
In each, the ε tolerance already existed — in a standard or a contract. No detector was trained; the codec's surprise is what points to the event.
The polarization of light in the Curie cable rotates with an earthquake. The tube holds 94% of the day; when the waves of the M7.4 Oaxaca quake arrive, the deadband collapses — at the exact time of the event.
Frequency (60 Hz in Brazil, 50 Hz in the UK) sits in a standards band. Over 6 real months, the tube holds 94.7%; every generator trip breaks the tube — the surprise spikes 14× above rest.
Models get stuck repeating themselves. Certified search (SIEVE, zero false negatives) catches all three loop types — literal, cyclic and paraphrased — with no false positives, where heuristics and plain compression have a blind spot.
One fiber becomes 101 sensors along the cable. The surprise detects the event with spatial resolution: the distance×time map lights up following the wavefront — quiet deadband 81% → event 31%.
The free detector works for strong broadband transients (earthquakes, blackouts, seismic waves). But a whale on a DAS cable or partial discharge on a power cable are narrowband, low signal-to-noise signals: the tube breaks everywhere and raw surprise doesn't isolate the event (we measured — correlation 0.11). They need dedicated band selection. Mapping where it wins and where it doesn't is part of the method.
In each domain, the tolerance wasn't invented by us: TVE ≤ 1% in the synchrophasor standard, the frequency band in the grid code, ±10 cm on an agent's position, half a cent in a freight contract. The ε already lives in a document — and it is exactly the radius of the tube.
Because the codec guarantees |error| ≤ ε on every sample, when it has to speak that means, provably, that the signal left the tolerated band. Detection inherits the guarantee from compression. It is the same family as the other pieces — the certified evidence plan (TUBE · TRUSS · CLAMP · SIEVE), where the rule is always the same: proof, not confidence.
If your signal has a tolerated band written down somewhere, the compressor can become the detector — free and certified. Bring the ε; we'll show you the surprise.