Yes, approved
No rule fires on this answer, for a third-party or fine-tuned model that reaches a consumer. A reviewer moves on.
Free snapshot · question 7 of 9
Is there an approved record of what that model was trained or tuned on?
A provenance record states where the training or tuning data came from, what period it covers, roughly how much of it there was, and what you took out before training, such as personal data or one customer's proprietary content. A commit message or a notebook is not a provenance record.
Why a carrier asks
What the answer decides.
For a model you fine-tuned or bought, this is the finding most likely to be raised. An examiner wants to know what went in, where it came from, and what was stripped out. Engineering notes are not the same as an approved record.
What the engine does with each answer
Derived, not written.
Each outcome below was produced by running Vaulith's rules engine on that answer, for a third-party or fine-tuned model that reaches a consumer. The same rules run in the paid assessment, on your whole estate, with the arithmetic shown.
Described somewhere, not approved
Will escalateR-PROVENANCE
No training-data provenance for the system you described
The training corpus is described in engineering documentation but has never been approved as a record. Because the model is consumer-impacting, it is the one an examiner is most likely to name, and the gap sits on exactly that model.
The examination items it touches
Where this lands in the questionnaire.
- AIS-3.4 For each high-risk model, describe the data used in training, including provenance, third-party sources, and any use of consumer data.NAIC AI Systems Evaluation Tool, training data provenance
- AIS-3.5 Describe the de-identification or minimisation applied to consumer data before it is used in training.NAIC AI model bulletin §4.2
- AIS-2.4 Provide current model documentation for each high-risk system, reflecting the version now in production.NAIC AI Systems Evaluation Tool, model documentation
Answer it about your own estate.
The snapshot asks this and eight more, in about ninety seconds, with nothing leaving your browser, and returns every finding with its reasoning and its remediation.