Clause 3.1, Stability
Stability is defined as the property that lets an AI system give the same output when its input is perturbed in ways the team has stated in advance. The subject is the whole system, the effect is an unchanged output, and the trigger is a named perturbation.
The requirement in plain English
Stability is defined as the property that lets an AI system give the same output when its input is perturbed in ways the team has stated in advance. The subject is the whole system, the effect is an unchanged output, and the trigger is a named perturbation.
What the public record says
The public view gives the term and one sentence. It adds no note and no example.
Why it matters to you
This is the one definition the FDIS makes public, and it has moved. Part 2 defined stability as the extent to which a neural network's output remains the same when its inputs change. The FDIS puts the whole system, not only the network, at the centre. It asks for an unchanged output rather than a degree of sameness, and it requires the perturbations to be named. A test built on the Part 2 wording can pass while the FDIS claim fails, because the system boundary and the perturbation set are different. If you brief engineers with the old definition, this page is the correction.
Evidence a reviewer asks for
- The perturbation specification: the dimensions changed, the ranges, the generator and its version.
- The system boundary, showing what is inside the AI system and what is outside.
- The invariance criterion: what counts as the same output, and at what tolerance.
- The test log with seed, draw count and result for each perturbation set.
Checks to run this month
- Is the subject of your stability test the AI system or only the model?
- Are the perturbations specified in writing before the run?
- Would a reviewer accept your definition of invariant output?
Dates
FDIS registered 2026-08-21. No date is attached to the term.
Cross-references
What this page does not state
The public text gives no note, example, tolerance or test method for stability.