11.3 Functional Performance Testing of Probabilistic MLS

From a dashboard number to a statistical claim

ML functional performance testing asks how well a model does the job it was hired to do. You still measure accuracy, recall, precision, and F1-score, and you still compare those numbers with acceptance criteria. Learning objective AI-6.1.3 (K2) is narrower than know the four metrics. It asks you to explain how this testing is carried out for probabilistic machine learning systems.

Probabilistic MLS do not give you a deterministic always 7 the way a conventional adder does. Two evaluation runs can differ. A single pass/fail flag on one lucky batch hides that uncertainty. Testing therefore moves beyond a binary status to a statistical measurement of performance against the criterion. You evaluate behavior across a large, representative set because one handful of examples cannot speak for the operational input domain.

A precondition is that the criterion itself is written in statistical terms. Instead of accuracy must be 98%, a serious requirement names three pieces:

• a performance metric (accuracy, recall, and so on),
• a margin of error (MoE),
• a confidence level (CL).

Those three together can be used to determine the minimum number of tests required. You need the concept of that relationship. You do not need to derive or compute sample-size formulae in the exam. The numbers later in this section are syllabus illustrations for understanding.

What MoE and CL are saying, without a z-table

Margin of error is the plus-or-minus band around a measured rate. If a report says 98% with MoE ±4%, the measurement is being presented as a band, not as a magical exact 98.000%. The band is about the precision of this test, not a promise that every future production day will land inside it.

Confidence level is how often that style of band would cover the true long-run rate if you repeated the whole measurement procedure. A 95% CL is a common engineering choice. It is not certainty. It is also not a coin flip. It answers how sure are we in this measurement method?, not how safe is the warehouse?

Sample size is the number of independent test cases you run. As that number grows, you have two planning options:

• Fix the CL and the MoE will decrease. The result becomes more precise (a tighter band).
• Fix the MoE and the CL will increase. The result becomes more certain (a higher confidence that the band method covers the truth).

That is the whole conceptual machine. More evidence either shrinks the band or raises your confidence in a band of the same width. Testers who only memorize a pass ratio skip this machine and then cannot read a statistically worded requirement.

Why does planning often look pessimistic? When you do not yet know the true accuracy, the most demanding ordinary assumption is that it could be anywhere from 0% to 100%. Variance of a proportion is largest near the middle, so that conservative planning assumption produces the largest uncertainty you still have to cover. It guarantees the requested MoE under a wide range of true rates. After you observe a very high accuracy, the achieved band is often tighter than the planned band, because a 98% process is less variable than a 50% process. You do not need the algebra; you need that qualitative story.

Syllabus illustration: 98% accuracy, ±4% MoE, 95% CL

The syllabus walks a single worked picture so the three knobs feel real. A criterion might be 98% accuracy with a MoE of ±4% at a 95% CL.

To confirm that the measured accuracy has a maximum MoE of ±4% at 95% CL, a sample size of 601 test cases is required under the conservative planning assumption above. To meet the 98% target on that 601-case set, at least 589 of the 601 cases must pass.

If you actually run all 601 and the observed accuracy is 98%, the measured MoE at 95% CL is narrower than ±4%—approximately ±1.1%—because variance is lower at high accuracy. The planned 601 was an insurance policy against not knowing the true rate in advance. The observed high rate pays you back with a tighter band.

You also do not always have to grind through all 601 cases. Sequential testing analyzes results as they accumulate and can stop early when the evidence already supports (or already rejects) the accuracy target. If observed accuracy stays consistently high (for example, no less than 98%), the required MoE of ±4% at 95% CL can be reached after about 170 test cases, and testing may conclude earlier.

Treat every figure in that paragraph as an illustration. Candidates are not required to derive or compute sample-size, MoE, or confidence-interval formulae in the exam. If a question asks what the 601 / 589 / 170 story is doing, answer in concepts: conservative planning sample, pass count that meets 98% of 601, tighter achieved MoE at high accuracy, and optional early stop when accuracy stays high.

Safety-critical illustration: 99% reliability with 95% confidence

Safety-critical systems sometimes use a stricter reliability-style requirement, such as 99% reliability with 95% confidence. In the syllabus illustration, that picture requires 299 test cases, all of which must pass. The shape differs from the 601-case accuracy example: you are not allowed a handful of failures inside a band; you are demonstrating a very high reliability claim with no failures in the sample. Again, this is an illustration for understanding, not an invitation to invent new formulae under time pressure.

The dataset you must use, and the sentence you must report

To validate these criteria you need a large test dataset that is completely independent of the training and validation datasets. Independence protects you from congratulating a model for memorizing data it has already seen. The set must also be a representative sample of the operational input domain. A statistically pretty interval on yesterday's easy cases is not evidence about tomorrow's messy cases.

Execute the cases with the model inside an ML development framework that can support statistical analysis. Then interpret the aggregated results with statistical confidence, not as a simple passed/failed ratio. A final report would say something like: the model achieved an accuracy of 94% ±4% at the 95% CL. Stakeholders can see the measured rate, the band, and the confidence recipe. They can then decide whether that functional performance is acceptable to deploy.

A bare 565 of 601 passed hides the band and the confidence. A bare passed hides everything. On the exam, prefer the reported sentence that still carries metric, MoE, and CL.

Putting the illustration numbers in a tester's pocket

Keep the following as landmarks, labeled in your memory as illustration only:

Notice what the landmarks are not. They are not a claim that every product in the world must use 601 cases. They are not a claim that you will be asked to compute 601 from a blank page. They show how a metric, an MoE, and a CL travel together, how a conservative plan can be larger than the sample you would need once accuracy is clearly high, and how a reliability demonstration can demand a clean sweep of a smaller set.

A worked decision, still without algebra

A hospital wants a triage ranking model. Product says we need 98% accuracy. A tester who stopped in Chapter 3 might run 50 charts, see 49 correct, and stamp Passed. A tester who absorbed this section asks three follow-up questions. What MoE is tolerable? What CL does clinical governance want? Is the evaluation set independent of training and actually like the night-shift intake mix?

If governance agrees to ±4% at 95% CL, the conservative illustration says you should be thinking in hundreds of cases (601 in the syllabus picture), not dozens. If early results sit at 98% or better, sequential logic may let you stop near 170 rather than burning the full 601. If the system is instead framed as 99% reliability with 95% confidence, the illustration flips to 299 all-pass cases. None of those choices requires you to look up a z-value. They require you to refuse a bare pass ratio as the acceptance language.

When you write the test report, copy the shape of the syllabus sentence: a measured accuracy of 94% ±4% at the 95% CL is a claim people can argue with. Looks good is not.