Purpose
Own lineage, quality, labeling, retrieval-index governance, and train/serve consistency for data that shapes AI behavior.
Engineering philosophy
Privacy asks who may see data; governance asks whether the data is fit for purpose. Bad corpora and poisoned indexes create silent production failures.
Why it matters
Hallucinations, biased outcomes, and retrieval failures often originate in unmanaged data—not in the model alone.
Common failures
- RAG indexes with no freshness or ownership
- Eval and fine-tune sets without label quality review
- Feedback loops that reinforce errors into memory or training
- No lineage from production answer to source document version
Mandatory checks
Gate controls. Each check is pass/fail via artifact + pass condition. Expected from the annotated capability level when the system meets the minimum criticality tier.
Production retrieval corpora and indexes shall have owners, versioning, and refresh cadence
- Artifact
- Corpus/index inventory with owner, version ID, and refresh schedule
- Pass condition
- 100% of production indexes have owner + version ID + refresh cadence; 0 indexes missing any field; stale beyond cadence flagged or rebuilt
Data used for evals or fine-tuning shall have documented provenance and quality criteria
- Artifact
- Dataset cards or registry entries for eval/fine-tune sets
- Pass condition
- 100% of datasets used in production gates or fine-tunes have provenance + quality criteria documented; promotion blocked if missing (CI or review checklist evidence)
Feedback or memory promotion paths that affect future answers shall be governed
- Artifact
- Promotion policy + write-path controls for feedback→memory/training
- Pass condition
- 100% of promotion paths require policy check or human approval; tests show ungated promotion to durable memory/training is denied
Recommended checks
Strengthen posture beyond the gate. Same measurable structure; non-blocking unless elevated by organizational policy.
Automated freshness and coverage metrics for critical corpora
- Artifact
- Corpus freshness dashboard (or job report) + alert config export covering each critical corpus ID
- Pass condition
- Each critical corpus has a documented freshness SLO (e.g. max age hours); ≥95% of sampled docs meet the SLO in the last 7 days; alert fires on freshness breach
Train/serve skew monitoring for features or embeddings
- Artifact
- Skew monitor config + last weekly skew report comparing train vs serve feature/embedding distributions
- Pass condition
- Skew job ran within the last 7 days for every production embedding/feature pipeline; at least one documented threshold exists; breach creates a tracked ticket or page
Dataset cards for major eval and fine-tune sets
- Artifact
- Dataset cards (or equivalent metadata) for each major eval and fine-tune corpus in the registry
- Pass condition
- 100% of major eval/fine-tune sets used in production promotion have a dataset card with purpose, source, PII handling, and last-updated date ≤12 months
Evidence required
- Corpus/index inventory with owners
- Data quality criteria and sample reviews
- Lineage examples from answer to source version
Severity & risk
- Severity
- high
- Impact if violated
- Risk level
- high
- Typical residual risk (impact × likelihood)
Engineering best practices
- Treat indexes as release artifacts with version IDs
- Separate experimental corpora from production indexes
- Review user-generated feedback before it becomes training signal
Automatic validations
- Index build pipelines emit version and checksum
- Alerts on stale critical corpora
- CI checks that eval datasets meet schema/quality gates
Manual validations
- Periodic corpus content audits
- Label quality sampling
Examples
- A policy RAG index is rebuilt nightly with a new version ID and canary before cutover
- Fine-tune data requires dual review before entering the training set
References
Crosswalks
MAP Map
NIST AI Risk Management Framework · supports
Fair Fair — Harmful Bias Managed
NIST AI Risk Management Framework · partial
§4 Context of the organization
ISO/IEC 42001 · aligns-with
Annex A AI system controls (selected themes)
ISO/IEC 42001 · partial
LLM04 Data and Model Poisoning
OWASP Top 10 for Large Language Model Applications · supports
LLM08 Vector and Embedding Weaknesses
OWASP Top 10 for Large Language Model Applications · supports
P-series Privacy (selected)
SOC 2 Trust Services Criteria · partial
Future evolution
Portable dataset and index manifests with integrity proofs across vendors.