Updated October 26, 2026.
Begin with the task, not the data label
An OCR test may need authentic handwriting. A loan-processing regression test may need hundreds of document sets with consistent borrower and property details. A new claims workflow may need examples of exceptions that have not yet appeared in production.
Each task needs different evidence. Whether the data is real, sanitized or synthetic is only part of the decision. The more useful questions are what must remain realistic, what must be controlled, and what the team is permitted to use.
Use reviewed real content when source detail matters
Real documents can preserve details that are difficult to anticipate in a template: annotations, corrections, unusual layouts, mixed handwriting and print, or the artifacts left by repeated scanning.
When those characteristics matter and the intended use is permitted, prepare a reviewed derivative. Define the sensitive information to remove, preserve the content needed for the task, and verify the exported result. Sanitization here means preparing real content under a specified policy; it is not a universal claim of anonymity.
Redaction can also change usefulness. Removing a date, identifier or passage may break a relationship or remove evidence the test depends on. Review both what remains exposed and what the downstream task can still measure. Do not assume that a redacted dataset preserves every statistical property of its source.
The GritWorks handwritten-clinical-data case study describes an enterprise pilot using reviewed source documents for controlled retrieval and agent testing. It illustrates why retaining real handwriting can matter; it does not establish that any redacted clinical corpus is suitable for any recipient.
Use synthetic content when control or missing coverage is the priority
Synthetic generation is useful when the team needs known values, repeatable variation, rare scenarios or examples before suitable production data is available. It can also be useful when real data exists but distributing it is unnecessary or inappropriate.
For a document workflow, realism includes more than plausible names. The layout must fit the task, required fields must be represented, and relationships must be deliberate. A borrower name that changes accidentally across a loan packet creates a different scenario from a consistent packet. Both may be useful, provided the expected result is explicit.
GritRender generates synthetic documents, images and audio from configured templates, fields and rules. Teams can use controlled variations to test difficult conditions. That capability should not be described as a blanket guarantee that generated data reproduces a production distribution.
Validate generated examples against the intended workflow. Keep a separate, approved evaluation set where feasible so that a system is not judged only on the same templates or patterns used to develop it.
Combine the approaches when both fidelity and variation matter
A hybrid workflow starts with source content that has been approved and reviewed for the intended use. The retained structure can become a template for new synthetic values and controlled variations.
GritWorks' published loan-document case study describes a production deployment using GritRedact and GritRender in this sequence. The workflow retained document layouts while generating consistent borrower, property and loan values across related files.
This is useful when generic templates miss important structure. It still requires checking the seed, the generated output and the rights to use both. Generation does not automatically remove information left in a source template.
Keep extraction separate from sanitization
Some workflows need structured fields rather than a document corpus. GritTransform extracts information from PDFs and images into JSON that an application can use. Review which fields are necessary and where they will go.
Extracted values can still be sensitive. PII flagging helps identify values for review; it does not remove them from the original document. Use GritRedact when policy requires content removal. Use GritRender when the evaluation needs controlled examples with known values.
A practical decision sequence
- Define the task: Specify what the model or application must do and how success will be checked.
- Confirm permitted use: Establish the recipient, purpose and environment before preparing source records.
- Identify essential realism: Decide which layouts, handwriting, relationships or recording characteristics must be retained.
- Identify missing coverage: List the rare, negative and boundary cases that the current corpus does not contain.
- Choose and validate the mix: Use reviewed real content, synthetic examples or both, and check privacy, usefulness and coverage separately.
Match deployment to your requirements
GritRedact, GritTransform and GritRender are available offline on-premises and through sales-assisted SaaS, with web-platform and customer API access after onboarding. Offline processing keeps processing within the chosen enterprise environment; hosted processing follows the agreed service and data-handling arrangements.
Bring a document type, a downstream task and the cases your current dataset cannot cover. We can help assess which preparation and generation workflow gives your team useful evidence.
