Independent and not affiliated with the FDA, MHRA, ISPE, PDA, or any agency. Get the appgoutham@madhadi.com
madhadi.comData Integrity & GxP Quality
Browse all topics → Articles Templates & Procedures Learning paths GlossaryScenariosToolsRegulatory ReferencesLearning PathsTopics About Start here
Checklist Plug-and-play starting point AI & Automation

Checklist: AI Training Data Readiness Release Gate

A plug-and-play sign-off checklist that operationalizes data readiness as a formal release gate before AI or machine learning modeling begins: profiling, contextualization, ground truth, ALCOA+ and lineage, split and leakage control, and bias and representativeness, each with pass, fail, or not-applicable disposition and a filled specimen.

Document type: Checklist

Read and copy the template below into your own quality system. It is a generic starting point for your own internal use, provided as is, with no warranty; see the Terms and License. Adopting it does not by itself create compliance.

This is a ready-to-use checklist. It is a control document, meant to be executed and signed, not a narrative to read once. Replace every <<FILL: ...>> placeholder with your own specifics. A worked filled specimen follows so you can see the level of detail an inspector expects. Verify each cited regulation against the current source before you rely on it.

Document control header

FieldEntry
Checklist titleAI Training Data Readiness Release Gate
Document number<<FILL: CHK-ID, e.g. CHK-AI-018>>
Version<<FILL: version>>
Effective date<<FILL: date>>
Owner<<FILL: role, e.g. Data Governance Lead>>
Governing procedure<<FILL: SOP-ID>> or AI training data readiness assessment SOP

Instructions

  1. Complete this checklist for every dataset before it is used to train, retrain, or run an AI/ML model that influences a GxP decision.
  2. Mark each item Pass, Fail, or N/A. Every Fail requires an entry in the “Evidence / notes” column describing the gap and the remediation, and blocks the gate until resolved or explicitly accepted per your deviation process.
  3. N/A requires a one-line justification; it is not a way to skip an item that actually applies.
  4. This checklist does not replace the readiness assessment record; it is the signoff gate that confirms the assessment was actually done and evidenced.
  5. Route the completed checklist to Quality Assurance for the final gate signature. Modeling does not begin until the gate is signed.

Section A: Intended use and sourcing

#ItemPass / Fail / N/AEvidence / notes
A1Intended use is defined, including whether the output is advisory or acts with less human review<<FILL>><<FILL>>
A2Every source system feeding the dataset is identified with a named data owner<<FILL>><<FILL>>
A3The extract query, parameters, and date are documented and reproducible<<FILL>><<FILL>>

Section B: Data quality profiling

#ItemPass / Fail / N/AEvidence / notes
B1The dataset is profiled for completeness, accuracy, consistency, timeliness, uniqueness, validity, representativeness, and traceability<<FILL>><<FILL>>
B2Acceptance thresholds for each dimension were set before profiling, not adjusted after seeing the result<<FILL>><<FILL>>
B3Units, formats, and categories are harmonized across sources; censored or qualified values (for example ”< 0.1”) are preserved, not coerced<<FILL>><<FILL>>
B4Numeric ranges were checked against the process specification, not only against the data’s own minimum and maximum<<FILL>><<FILL>>
B5Every profiling gap has a documented remediation or an explicitly accepted, justified limitation<<FILL>><<FILL>>

Section C: Contextualization and features

#ItemPass / Fail / N/AEvidence / notes
C1Raw values are anchored to batch, phase, or process context, not used as bare numbers<<FILL>><<FILL>>
C2Every engineered feature has a written definition, source, and reproducible transformation<<FILL>><<FILL>>
C3The feature logic used to build the dataset is confirmed identical to the logic that will run at inference time<<FILL>><<FILL>>

Section D: Ground truth and labeling

#ItemPass / Fail / N/AEvidence / notes
D1The ground-truth source is documented: existing structured record, expert labeling, or objective downstream outcome<<FILL>><<FILL>>
D2Where labeling was used, it followed a documented labeling procedure with qualified labelers<<FILL>><<FILL>>
D3Inter-rater agreement was measured on a double-labeled subset and meets the pre-set threshold<<FILL>><<FILL>>
D4Disagreements were adjudicated by a defined method and the resolution is recorded<<FILL>><<FILL>>
D5The label set is versioned<<FILL>><<FILL>>

Section E: ALCOA+ and lineage

#ItemPass / Fail / N/AEvidence / notes
E1The source data remains in its governed system; the extract is a documented true copy<<FILL>><<FILL>>
E2Every transformation is captured as version-controlled, re-runnable code, not manual edits<<FILL>><<FILL>>
E3Every excluded record has a documented reason on the lineage and exclusion log<<FILL>><<FILL>>
E4The final dataset is frozen, versioned or hashed, and retained per the records retention schedule<<FILL>><<FILL>>

Section F: Split and leakage control

#ItemPass / Fail / N/AEvidence / notes
F1The grouping unit (batch, lot, or equivalent) and time structure of the data are identified<<FILL>><<FILL>>
F2The split is by group, so no single group appears in both training and test<<FILL>><<FILL>>
F3The split is time-based where the model will run forward in time in production<<FILL>><<FILL>>
F4The test set is locked, version-controlled, and used only once<<FILL>><<FILL>>
F5All preprocessing (scaling, imputation, feature selection) was fit on training data only<<FILL>><<FILL>>
F6Every feature was audited for availability at the moment of prediction; none encodes the outcome<<FILL>><<FILL>>

Section G: Bias and representativeness

#ItemPass / Fail / N/AEvidence / notes
G1Coverage of the dataset against the production population (site, product, instrument, regime) is tabulated<<FILL>><<FILL>>
G2Class balance for the target variable is quantified and evaluation metrics were chosen to survive imbalance<<FILL>><<FILL>>
G3Every under-represented subgroup that matters to the intended use is either remediated or explicitly excluded from scope, with the exclusion documented<<FILL>><<FILL>>

Gate disposition

FieldEntry
Total items<<FILL: count>>
Pass<<FILL: count>>
Fail (open)<<FILL: count>>
N/A<<FILL: count>>
Overall gate outcome<<FILL: Open for modeling / Blocked pending remediation>>
Prepared by (name, signature, date)<<FILL>>
QA gate signature (name, signature, date)<<FILL>>

References

21 CFR Part 11 and EU GMP Annex 11, for the electronic-record controls the dataset and its lineage must satisfy. FDA, Data Integrity and Compliance With Drug CGMP, Questions and Answers (Guidance for Industry, December 2018), for the completeness, consistency, and accuracy expectations this checklist operationalizes. ICH Q9, Quality Risk Management, for sizing the gate’s rigor to the model’s intended use and consequence. MHRA GxP Data Integrity Guidance and Definitions; PIC/S PI 041 (reference by title; describe, do not paste).

Confirm the current version and clause numbers of each reference before issue.


Filled specimen

The following shows the checklist completed for an illustrative bioreactor yield-prediction model. Numbers and findings are illustrative; replace with your own.

#ItemResultEvidence / notes
A1Intended use definedPassAdvisory yield forecast at hold-decision points; process engineer confirms before any hold/release action
B1Quality dimensions profiledPassProfiling report DATA-PROF-2026-041; dissolved-oxygen 4% missing overall
B3Units and qualifiers harmonizedFail, resolvedInitial extract mixed C and K units on one tag; corrected in pipeline v1.2, re-profiled Pass
C3Feature logic identical train/inferencePassFeature spec FEAT-SPEC-YIELD-v2 deployed in both the training pipeline and the inference service from the same repository tag
D1Ground-truth source documentedPassObjective outcome: final released yield from batch disposition record, not a subjective label
E3Exclusions documentedPass6 batches excluded for corrupted sensor logs, logged in LOG-AI-YIELD-2026-07 with root cause
F2Split by groupPassSplit by batch ID; verified no batch ID present in both train and test partitions
G3Under-represented subgroup dispositionFail, remediation plannedLine 3 is 6% of records (new line, limited history); model intended use scoped to exclude Line 3 until 6 months of additional data collected, documented in readiness assessment record RA-YIELD-2026-08
Overall gate outcomeBlocked pending remediation, thenOpen for modeling with Line 3 excluded from intended use, after QA review of the scope statement

This specimen shows the checklist doing its job: two genuine Fail findings, both resolved or explicitly scoped rather than waved through, with the evidence column pointing to the actual record instead of a bare checkmark.

Common inspection findings this checklist prevents

  • A checklist exists but every item is marked Pass with no evidence reference, so the review cannot be verified.
  • Modeling started before the gate was signed, or the gate was signed after the model was already built.
  • A Fail item was overridden with no documented remediation or acceptance rationale.
  • N/A was used to skip an item that genuinely applied to the dataset.
  • The checklist was completed once and never re-run when the dataset was later re-extracted or the model retrained.

How to adapt this checklist

  1. Set your document number, owner, and the governing SOP reference in the header.
  2. Adjust the sections to match your organization’s actual readiness procedure; add or remove items only if the underlying control genuinely differs, not to make the gate easier to pass.
  3. Decide your escalation path for a Fail: who can accept a documented, scoped limitation versus who must require full remediation before the gate opens.
  4. If you build both tabular models and retrieval-based generative assistants, issue a parallel gate for the document-corpus pattern (freshness and versioning, redaction, chunking, retrieval evaluation, citation) rather than forcing this checklist’s tabular-specific items onto that different kind of dataset.
  5. Confirm every regulation in the references section against the current published version before issue.
Use madhadi.com as an app Full screen, works offline, one tap from your home screen.