This is a ready-to-use record for the integrity of a GxP AI/ML model’s training data. The integrity of a supervised model is bounded by the integrity of the data it learned from, so this record is a GxP record in its own right: it lets a future investigator reproduce the model and lets an inspector trace the chain from data to decision. Replace every <<FILL: ...>> placeholder with your own specifics. A worked filled specimen follows. Verify each cited regulation against the current source before you rely on it.
1. Source and lineage
2. Representativeness
State honestly what the data does not cover. A model trained only on routine cases will fail on the rare events that matter most, and an unstated gap is a finding.
3. Labeling quality (supervised models)
4. Class balance and splits
5. Retraining data provenance (if applicable)
Data pulled from production to retrain a deployed model is itself GxP data and carries the full ALCOA+ expectations. Retraining is a data-integrity event, not just an engineering task.
6. Acceptance criteria for this record
- Source, extract date, record count, and lineage are stated and reproducible.
- Representativeness is described, including honest statement of gaps.
- For supervised models, the labeling SOP, labeler qualifications, inter-rater agreement, and disagreement resolution are recorded.
- Class balance is known and the metrics chosen are not fooled by imbalance.
- The train/validation/test split is defined, the test set is locked and version-controlled, and a leakage check is recorded.
- The frozen dataset is versioned so the model can be rebuilt.
- Any production data used for retraining carries ALCOA+ controls and a change-control reference.
7. References
21 CFR Part 11 and EU GMP Annex 11 for electronic records around the dataset.
MHRA GxP Data Integrity Guidance and Definitions; PIC/S PI 041 (reference by title; describe, do not paste).
ALCOA+ data-integrity principles (attributable, legible, contemporaneous, original, accurate, plus complete, consistent, enduring, available).
GAMP 5 Second Edition (ISPE, 2022) for the AI/ML lifecycle context (reference by title).
Track the draft EU GMP Annex 22 and Annex 11 revision (2025); confirm status before relying on either.
Confirm the current version and clause numbers of each reference before issue.
8. Revision history
9. Approvals
Filled specimen
The following shows the record completed for an illustrative complaint-classification model. Numbers are illustrative.
This package lets a future investigator rebuild the exact model and shows an inspector the chain from source data to the labels to the frozen set. The honest note about under-represented combination-product complaints is exactly the kind of stated gap that reads as a controlled program rather than a hidden weakness.
Common inspection findings this record prevents
- A model that cannot be reproduced because the training dataset was never versioned.
- Labels generated with no documented SOP or labeler qualifications.
- No inter-rater agreement for subjective labels, so label consistency is unknown.
- A hidden data gap that surfaces as production failures on an under-represented population.
- Production data used to retrain with none of the ALCOA+ controls that GxP data requires.
How to adapt this record
- Fill the source, extract, and lineage so the dataset can be rebuilt from scratch.
- State representativeness honestly, including what the data under-represents.
- Record the labeling SOP, labeler qualifications, inter-rater agreement, and how disagreements were resolved.
- Define the splits, lock the test set, version the frozen dataset, and record a leakage check.
- If you retrain on production data, attach the ALCOA+ controls and the change-control reference.
