Two values lie at the boundary.
Every measurement has an error margin. If the contract only knows "meets" and "fails", a result near the limit gets one of the two names, and which one depends on the luck of that measurement. Measure again and it may come out the other way.
This is not a hypothesis. In one of our studies, changing only the seed of the random draw used to compute the margin, with no new data, changed 3.6% of the verdicts that looked decided. None of them had been decided by the measurement; they had been decided by the draw.
Near the line, whoever picks which measurement goes into the report picks the verdict. The third value takes that choice off the table: when the error band crosses the limit, the contract already says what to do, and nobody has to fight over the name.
Six things the clause has to say.
The limit, without ambiguity
Does "up to 200 ms" include 200? Does "from R$ 199" include 199? In one notebook case, code read "from" as "above" and charged shipping to people it should not have. Write "at most" or "at least", with the number.
The method's error margin
How far the measurement can be off in each direction, declared by the method before measuring. A margin that only appears after the result is negotiation, not measurement.
The margin of safety
How many times the error margin the result must be away from the limit to decide. Once is the minimum; more protects you when the margin itself is uncertain.
Who measures, with what data
A third party, the buyer, or the supplier with the data open. If whoever is measured measures alone and keeps the data closed, the clause is worth whatever they want.
When the verdict is declared
With what data, in what window, with a timestamp. The result comes out together with the margin and margin of safety used, so anyone can redo the math.
What to do when undecidable
Measure again, with more samples, on a short deadline, with who pays already defined. Meanwhile, no penalty and no acceptance. This is the clause that prevents the dispute.
A margin chosen afterwards is cheating under another name.
The temptation is to leave the margin of safety open and pick it when the result arrives. Whoever is losing always finds a margin that saves them. That is why it has to be in the contract before the first measurement.
We measured this in one of our tools, the sensor attestation: requiring a margin of 1.5 times the error changed nothing; requiring 2 times wiped out the verdicts that flipped by chance. Picking 2 because it gave the pretty result would be exactly the mistake. We tried three ways of reaching the number without looking at the result: two gave similar values (1.4 and 2.0) and the third pointed to no value at all.
The practical lesson: choose the margin by the cost of being wrong, not by the result. If erring in the supplier's favor is expensive, use a bigger margin on that side. And write the number down.
The same structure, five contracts.
| Contract | What is measured | Where the margin comes from | What to do when undecidable |
|---|---|---|---|
| Response time | typical time of an operation, in milliseconds | the variation between repeated runs | a new round with more runs, at an agreed time |
| Availability | share of the month the service responded | probe frequency and measurement failures | cross-check both sides' logs before applying a reduction |
| Works or equipment | an acceptance quantity, such as concrete strength or antenna performance | instrument precision and number of samples | more samples from the same batch, by a lab chosen by both sides |
| AI model | accuracy on an agreed test set | the set's size and which cases were drawn | a bigger set, chosen by the buyer and never seen by the supplier |
| Pay-per-use sensor | whether the sensor is alive, from the size of the data it sends | the natural noise the manufacturer declares | the period is not billable, and the sensor is checked at the next visit |
The sensor case comes from a notebook entry: a live sensor never sends data smaller than its own noise, and a frozen one sends almost nothing. Judge what leaks, not what is declared.
Undecidable can be large.
An honest report can come out with a lot of undecidable. In one of our studies on program performance, of the differences a statistical test called real, 81.6% did not support a decision against the agreed limit (between 78.9% and 84.2%, depending on the draw).
In another case, an antenna report, only 0.18% of the points sat near the line, because the antennas were, at the median, 7.14 dB from the limit (on a synthetic sweep). The difference is not the method: it is where the cases live.
If undecidable is frequent in your contract, that is information: either the measurement is too imprecise for the chosen limit, or the limit sits right on top of normal operation. Both are fixed before signing, not after the first dispute.
Where this can mislead.
This is not legal advice
The generator builds a structure, not a text to sign. Every contract needs review by whoever answers for it, under the laws and standards of its sector.
The margin can be declared too small
Whoever declares a margin smaller than the real one makes everything decidable, and decides by chance. The margin has to come from the method, with a way to check that it is honest.
Undecidable can become a hiding place
A supplier with poor measurement can stay forever in the middle band. Cap how many undecidables in a row are accepted before measurement moves to a third party.
Who measures changes everything
With the data closed in the hands of whoever is measured, no three-value clause protects you. Open data, or a third party, comes before the margin of safety.
For use in practice.
Meets, fails and undecidable. The third says where the measurement stopped deciding. →
How many times the error margin the result must be away from the line to decide. Declared beforehand. →
Whatever decides whether something passed. Here, the measurement method with its margin. →
Does this apply to your case?
Tell us in two lines what you need to decide or measure. The first conversation is to see whether measurement solves your case, and if it does not, we say so.
← Verification · stickybit.com.br
- Our own measurements (2026): the 200-seed audit on the Go garbage collector study (3.6% of decided verdicts; 81.6% with a range of 78.9% to 84.2%); the same audit on the antenna report (0.18%, synthetic sweep, median margin of 7.14 dB) and on sensor attestation (margins of 1.5 and 2 times); three ways to derive the margin. Methods in The draw made the call.
- The "from R$ 199" limit case: Green is silence, not a certificate. The pay-per-use sensor: Judge what leaks, not what is declared.
- The clause generator runs in your browser; nothing you type leaves the page.