Clause 7, Statistical measures of robust behaviour

The public titles name five measure families beside the general clause: performance drops, the system's response to a targeted perturbation, separability metrics, statistical reliability engineering, and curvature analysis. The body text is withheld. A buyer can use the titles to choose a family and to check that the claim matches the measure.

The requirement in plain English

The public titles name five measure families beside the general clause: performance drops, the system's response to a targeted perturbation, separability metrics, statistical reliability engineering, and curvature analysis. The body text is withheld. A buyer can use the titles to choose a family and to check that the claim matches the measure.

What the public record says

The public table of contents lists clause 7 and subclauses 7.1 General, 7.2 Robustness assessment based on performance drops, 7.3 Targeted assessment of the system's response to a perturbation, 7.4 Assessment through separability metrics, 7.5 Assessment through statistical reliability engineering and 7.6 Robustness assessment through curvature analysis. The body text is not public.

Why it matters to you

Five families produce five different kinds of evidence. A performance drop is a before-and-after comparison and needs a clean baseline. A targeted perturbation asks how the system responds to one named change. A separability metric concerns how well classes or states remain distinct. Reliability engineering brings a failure and lifetime vocabulary. Curvature analysis is a geometric property of the model surface. A pack that reports a performance drop and calls it a robustness guarantee has crossed families without saying so. Naming the family is the cheapest defence against that.

Evidence a reviewer asks for

• The measure selection note, naming the family and the reason it fits the claim.
• The metric definition and the clean baseline.
• The result with its uncertainty and the sample it came from.
• The claim the measure can and cannot support.
• The reviewer sign-off on the match between measure and claim.

Checks to run this month

1. Which of the five families does the result come from?
2. Does the claim use a word from a different family?
3. Is the baseline measured on the same data and metric?

Dates

FDIS registered 2026-08-21. The clause body is not public.

Cross-references

• Set the acceptance threshold and the decision rule
• Uncertainty, complexity and model confidence

What this page does not state

The body text of clause 7 and its six subclauses is not public.