The butcher's scale carries a certification seal.
When you buy meat, you don't trust the butcher: you trust the scale. Not because it is good, but because it carries a calibration seal that an inspector can check at any time, without asking the owner's permission.
Sensor data rarely has that. A power utility, a financial market operator or a factory stores millions of measurements a day, and when an outsider asks "is this number right?", the answer is usually "trust us".
The evidence plane is the scale with a seal. It does not replace the customer's database or send raw data outside. Next to the normal storage, it produces a file with five guarantees, and hands whoever needs to audit an open verifier to check everything on their own.
It works with their system, not instead of it.
It sits beside, not in the middle. The plane runs inside the customer's network, in their private cloud or on their own servers, and copies the stream that already flows through their messaging system or the plant historian. If it stops, nothing on the main path stops. The usual storage keeps working; the evidence file comes out alongside.
The tolerance already exists in a document. We don't have to convince anyone how much each measurement may vary: that is already in the standard, the contract or the service-level agreement. For example, the standard for phasor meters on the power grid (IEEE C37.118) allows at most 1% error. The customer brings the tolerance; we guarantee it is met.
The verifier is open. A guarantee that requires "trust us" would betray the whole idea. The verifier is an open-source program that the regulator, the auditor or the other party to the contract runs on their own to check seals and tolerances.
Each guarantee is a tool, with the number it has already proven.
What was measured against an installed competitor comes with a number. The seal is a well-known building block (digital signature and a tree of fingerprints), with no performance number to show.
| Guarantee | Tool | What it does, plainly | Measured number | In practice |
|---|---|---|---|---|
| Faithful within tolerance | TUBE | stores a smaller file in which no measurement drifts from the original by more than the tolerance | 29× (GNSS) · 38.9× (DAS fiber) | years of data fit in a fraction of the disk, keeping the point-by-point guarantee |
| Sealed against tampering | GIRDER | stamps each block with a digital signature and links them all in a tree of fingerprints | well-known construction | changing a single number breaks the seal, and anyone notices |
| Answerable without opening | CLAMP | answers averages and totals straight from the compressed file, with a range that contains the exact answer | 187 of 187 days inside the range | the auditor gets the weekly average without decompressing years of data |
| Complete in search | SIEVE | finds every occurrence of a pattern, leaving none out, reading only what it needs from disk | 228× less disk reading | common indexes miss between 3% and 44% of the true results without warning; this one misses none |
| Watched | TRUSS | compact counting summaries that cannot be fooled by someone sending crafted data | Redis counted 1 of 30,000 | under attack, the ordinary count hid almost everything; the robust summary didn't |
Event detection comes as a bonus: when the stream leaves the tolerance, the compressor itself gives it away. That is certified surprise.
The plane became operating infrastructure.
On the same tolerance rule we built the pieces that were missing between the sensor and the audit. Each was measured on real data (motion sensors, robot arms, a car's data bus, energy meters, depth video, a self-driving car's laser sensor), with the experiment kept next to the code. And one of them failed, which is also part of it.
When the network chokes, an ordinary queue drops measurements and leaves holes with unknown error. This one widens the tolerance only as long as needed, declares the new limit and honors it.
delivered 100% vs 41% for the ordinary queueFixed memory that never forgets, only blurs: the past gets less sharp, with the blur recorded, and an incident stays at full sharpness, even if power fails.
24 KiB for 6 sensorsOld data gets more compact without the lie of averaging: every question returns a range that certainly contains the truth.
6,000 of 6,000 windows contain the truthMatching two sensors by clock is guesswork. On a real car's data bus, matching by nearest timestamp got 43.8% of pairs wrong without warning. Here each pair is classified as certain, ambiguous or impossible.
254,633 of 254,633 correct pairs containedThe sensor's physics says what could have happened in a data gap. And when it says "impossible", it found a defect: in a public robot-training dataset, the 39 violations were the 39 splices between recordings.
39 of 39 splices, no false alarmEach device sends only a compact index, and the cloud answers fleet totals with a guaranteed range. Raw data never leaves the edge.
2,000 of 2,000 ranges correctImages and depth video with per-pixel tolerance, and questions by region ("did the gripper stay within 5 mm?") without opening the file.
1.24 to 1.33× better than SZ3 on videoLaser data with per-coordinate tolerance and proof that a safety zone is empty, with no false negative by construction.
1,800 of 1,800 queries, zero false negativesSending different layers to each subscriber: built, measured, and plain delivery won on every measure. That record sits on the first line of the project.
the simple one wonFive signs separate obligation from convenience.
- The tolerance is already in a document: a standard, contract or service agreement.
- The buyer is compliance, not performance.
- The current vendor gives trust, not proof: "it's right because the system says so".
- Lots of data, kept for years.
- An outsider needs to check: regulator, auditor, counterparty.
When four or five apply, the platform stops being a convenience and becomes an obligation.
| Field | The tolerance comes from | Status |
|---|---|---|
| AI platforms (usage billing, search, audit) | service agreement, quota | measured |
| Games and match replays | game tick | measured |
| Weather and geographic data | field tolerance | measured |
| Market surveillance | order-record rules | thesis |
| Power grid (synchrophasor) | error ≤ 1% (IEEE C37.118) | thesis |
| Robotics | sensor tolerance | staged plan |
| Cold chain, meters, dams, batteries | health and safety rules | thesis |
| Freight | half a cent per quote | thesis |
Start with what is already measured.
AI platforms first
The only field already measured against the installed competitor (Redis and hnswlib). The pain is current and it validates three guarantees at once.
Financial markets as the story
The only place where the whole platform becomes a legal requirement. Good as a narrative, but a thesis until a pilot with real order data.
Games and the grid as cheap proofs
One is already measured (match replays); the other has the tolerance written in the standard.
Turn down the wrong vertical
Where the only requirement is in-memory speed, with no audit, the platform loses. That is focus, not a flaw.
Data that comes with proof that it is right and untouched, checkable by someone who doesn't trust the producer.
The signed fingerprint of each block. Change one number and the seal breaks.
The program that checks everything, with open code, for the auditor to run on their own machine.
Where this could be wrong.
Pure in-memory speed loses
With no need for completeness, long retention, adversaries or auditors, a simple in-memory search wins. We have measured that.
Several fields are still a thesis
Financial markets, the power grid and freight are mapped, but without a pilot on real customer data.
The seal doesn't say the sensor measured right
It proves the data didn't change after it was written. A miscalibrated sensor records wrong values with a perfect seal.
Numbers per tool, not per customer
Each number comes from an experiment on public or in-house data. On a customer's stream, compression and savings may differ.
← Certified telemetry · stickybit.com.br
- In-house experiments, 2026, with code and results kept in each project: compression (smartphone GNSS, DASCore DAS fiber), queries on compressed data (187 test days), complete search (corpus audit and hnswlib measurement), robust summaries against Redis 8.8, and the operating structures (queue, black box, archive, joining, gaps, fleet, image, point cloud, layered delivery).
- Standard for phasor meters on the grid: IEEE C37.118 (total vector error ≤ 1%).