Audit Data Hygiene

Purpose

Produce a field-level, read-only audit with traceable evidence and review actions. Preserve every supplied value and control state.

Required inputs

• Supplied records with stable record identifiers and source locations
• Field definitions, required fields, and allowed values, if available
• Evidence sources with verification dates
• Freshness thresholds or review policy, if available

Label absent inputs Missing; never retrieve or infer them.

Workflow

1. State the supplied datasets, scope, snapshot date, and policies. Do not imply access to a CRM or other system.
2. Inventory each record and field without normalization. Preserve disqualifiers, opt-outs, consent restrictions, and replies exactly.
3. Classify each issue as Missing (safe required value absent), Stale (threshold exceeded), Conflicting (sources disagree), Duplicate candidate (possible same entity), Unsupported (value lacks support), or Unsafe requirement (protected-trait or unnecessary sensitive-data field). Without a freshness rule report Freshness unknown, not Stale.
4. For each issue, name record, field, observed value or absence, and sources and dates. Show every conflicting value without choosing truth. Show both duplicate-candidate IDs and matching evidence; never merge or delete.
5. Assign severity: Critical for a violated opt-out or control state; High when eligibility, routing, consent, or identity could change; Medium when workflow reliability is impaired; Low for formatting or non-blocking completeness. Explain the impact without inventing one.
6. Propose one unexecuted human-review action per issue, such as source verification or conflict adjudication. Do not correct, enrich, normalize, merge, delete, or update records.
7. Reconcile issue counts by category and severity. List checked fields with no issue separately; absence of a detected issue is not proof of correctness.

Output format

# Data-hygiene audit draft
- Scope / snapshot / policy: [supplied values or Missing]
- Status: DRAFT ONLY — READ-ONLY AUDIT
## Findings
| ID | Category | Severity | Record | Field | Observed value(s) | Source(s) and date(s) | Reason | Proposed review action |
|---|---|---|---|---|---|---|---|---|
## Counts
- By category: [counts]
- By severity: [counts]
## No issue detected
- [record — fields checked]
## Unresolved inputs
- [missing policy, source, definition, or date]
## Actions taken
- None

Guardrails

• Never invent evidence, silently select an authoritative value, or claim freshness without a supplied threshold.
• Reject protected-trait targeting, segmentation, or decisions and unnecessary sensitive-data use. Classify such field requirements Unsafe requirement, not Missing; propose privacy/legal schema review, never collection or remediation.
• Never mutate, correct, enrich, normalize, merge, delete, suppress, contact, browse, or change any external system or system state.

Quality check

Confirm every finding names a record, field, sources, severity, reason, and review action; categories remain distinct; control states are preserved; counts reconcile; and actions taken are None.

Example invocation

Use audit-data-hygiene on the supplied fixture records; flag unsupported employer evidence and preserve opt-outs without correcting any system.