> ## Documentation Index
> Fetch the complete documentation index at: https://labs.laer.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Best practices for PII/PHI Review workflows

Use these best practices to improve detection results, produce defensible counts, and reduce rework across matters.

### Agree the element taxonomy before you run it

* Define what is reportable for this matter before configuration begins.
* Decide how expired identifiers, employee data, and residency limits are treated.
* Confirm whether clinical categories are in scope, not just health identifiers.
* Record matter-wide scope decisions in the general instruction and element-specific rules in the element definitions.

Scope decisions belong in the configuration. Applying them as manual corrections afterwards is slower and harder to defend.

### Start with clean inputs

* Use a dataset that matches the review scope.
* Verify that documents contain usable extracted text.
* Identify image-only and scanned material and process it before review, not after.
* Remove obvious duplicates before running the workflow.

Documents without text cannot be reviewed. In breach populations they are often the documents most likely to carry identifiers.

### Test on a sample first

* Run PII Detection on a pilot set before using the full dataset.
* Check the determination split against what you expect for this population.
* Read the supporting text behind a representative mix of element hits.
* Check the exception report early to confirm text extraction coverage.

A definition error found on a sample costs one run. Found on the full population, it costs two.

### Write definitions that say what does not qualify

* State exclusions explicitly. The most reliable definitions rule things out as well as in.
* Be specific about elements that only count in pairs, such as a card number with its security code, or a username with its password.
* Keep element names and tag names aligned when you rename either.
* Change one definition at a time and test before making further adjustments.

An element that fires on nearly every document is almost always a definition that is too broad.

### Review output at each stage

* Validate the document population before extracting individuals.
* Sample by confidence level rather than at random — the mid-range bands are where definitions are wrong.
* Review the documents where the determination and the model score disagree.
* Look at which elements are driving the population before treating the counts as final.

Early review prevents errors from carrying into the individual population.

### Verify the affected-individual count before reporting it

* Check for the same person appearing under two spellings, which shows up as near-identical names with different document sets.
* Check for two people consolidated under one common name, which shows up as conflicting addresses on a single record.
* Read the source documents for individuals identified from a single element in a single document.
* Verify the individuals with the highest element counts by hand — these drive notification obligations.

The individual count is the number that leaves the building. It deserves a deliberate QC pass.

### Reconcile the exception report

* Account for every document submitted, not just the ones that produced results.
* Re-OCR and re-run text extraction failures rather than accepting the gap.
* Record the reason for any exception that cannot be resolved.

A review that reports on most of a population and cannot explain the remainder has a gap someone will ask about.

### Use reviewer feedback to improve results

* Apply reviewer tags to a sample. They override predictions, drive re-evaluation of conflicting documents, and are what the metrics measure against.
* Refine definitions when the same correction appears repeatedly across documents.
* Run the metrics analysis against reviewed documents rather than judging a run by its overall counts.
* Recheck edge cases after making changes.

### Reuse what works

* Save successful configurations for similar matters.
* Keep a standard element taxonomy for common incident types and adapt it per matter.
* Compare results across matters to spot recurring issues.
* Keep internal notes on validated definitions and thresholds.