This is the vocabulary for the model itself: classification and regression models, generalisation, and the two failure modes, overfitting and underfitting.
What the base framework establishes
The group defines a classification model as one whose expected output is one or more classes, and a regression model as one whose output is a continuous value. It defines generalisation as the ability to make correct predictions on data the model has not seen, and then defines the two ways that ability fails. Overfitting is fitting the training data so closely that new data is handled badly; it can come from irrelevant features, noise or a model that is too complex for the data. Underfitting is the opposite: the model is too simple or too poorly specified to learn the pattern. The notes add the practical test, which is the gap between error on training data and error on separate validation data.
Why it matters to an AI management system
These four words decide whether an evaluation report is credible. A model card that claims high accuracy without saying whether the figure came from training or unseen data is using the vocabulary badly. The clause gives a governance lead the language to reject that card politely and ask for the right number.
The public draft summary names no change to 3.1 Model development and use. The final text could still touch it.
Evidence a reviewer samples
| Category | Artefact |
|---|---|
| Model card | The model type and the task it serves, with the evaluation dataset named. |
| Evaluation report | Training, validation and test results reported separately. |
| Overfitting decision | A note on the gap between training and production performance and what was done about it. |
One question to prepare for. Is the headline accuracy a training figure or an unseen-data figure?
Common gaps
- One accuracy figure is reported with no statement of the dataset it came from.
- The model card names an algorithm but not the task type, so the reader cannot tell classification from regression.
- A production drift signal is absent even where the training and production data differ.
Prepare before publication
Rewrite the model card so that task type, model type and the three dataset results are separate fields. Add one line on the overfitting check.
Dates that matter
No separate date attaches to 3.1; the record dates are the stage date, the planning estimate and the CEN expectation stated above.
Cross-references
6.2 Task · 6.3 Model · 8.5 Verification and validation
Source
Source for 3.1: ISO/IEC 23053:2022, clause 3.1. The statement is read from the public base sample at 3.1 and is written in our words. The two official records are linked on this page so any line can be checked.
Not legal advice; does not certify. A reading of the framework and the evidence a reviewer commonly samples, not a substitute for the standard or your auditor's rules.