Clinical validation

How BastionGPT Is Clinically Tested

Every note and document type BastionGPT supports goes through a structured validation program before we describe it as trained, tuned, and clinically tested. This page explains what that program covers, how each document type is tested, and where human judgment stays in charge.

99.3%
Sentence-level accuracy on clinical documentation tasks in a peer-reviewed study by Baylor College of Medicine and Texas Children's Hospital researchers.

What we validate

BastionGPT's validation catalog covers 100+ general and behavioral health document types across the workflows clinicians actually run: session and progress notes (SOAP, DAP, BIRP, and more), intake and diagnostic assessments, risk and safety documentation, treatment plans, psychological testing reports, care coordination and discharge records, consent and privacy forms, and the billing, school, disability, and legal documents practices are asked to produce. Coverage spans United States, Canadian, and Australian documentation expectations.

You can explore these formats in our free note and document template library, which pairs each template with a section-by-section structure guide and a complete sample note.

How we test each document type

Each type is exercised across five task classes, matched to how that document is used in practice:

Drafting

Generate a complete note or report from bullet points, a dictation, or a session transcript, at the length and structure the document type calls for.

Transformation

Convert between formats, such as DAP to SOAP or a narrative note to a payer-facing progress note, without losing clinically important content.

Extraction

Pull structured facts such as dates, diagnoses, scores, risk status, and follow-ups into fields exactly. A wrong name or date in a record is never a small error.

Privacy and segmentation

Redact protected information correctly, keep psychotherapy notes separate from the standard record, and keep each person's content clean in group and family documentation.

Administrative fit

Support billing documentation accurately, rewrite for patient, school, insurer, or legal audiences, and flag tasks that need a human decision instead of guessing.

What we measure

Each document type is scored against the failure modes that matter for that type rather than a single generic quality score:

Accuracy. Names, dates, scores, diagnoses, and service facts must match the source exactly.
Completeness. Every required section appears, with no missing risk or follow-up fields.
Clinical safety. Missed or invented risk indicators are treated as automatic failures.
Privacy. Redaction correctness, psychotherapy-note segregation, and release-of-information scope.
Billing support. Service, time, diagnosis, and medical-necessity linkage without unsupported codes.
Legal defensibility. Neutral tone, clear chronology, and appropriate expression of uncertainty.
Interoperability. Output keeps structure that maps cleanly into EHR templates and exchange formats.
Cultural responsiveness. The client's own explanatory model and context are preserved, in audience-appropriate language.

Stricter gates for high-risk documents

Suicide and violence risk assessments, safety plans, mandated reports, and legal documents are held to our strictest acceptance gates. Fabricated or omitted risk content is an automatic failure and legal documents are tested for scope discipline: neutral language, accurate chronology, and clear boundaries around what the record can support.

Human review by design

BastionGPT is built as documentation support for licensed professionals. It drafts, converts, extracts, and quality-checks; you review and sign. That design mirrors what record-keeping guidance and payer rules actually reward: accuracy, completeness, privacy discipline, and supportable reasoning, with clinical judgment as the final step.

Validation results and the clinical content we publish are reviewed by the BastionGPT Clinical Advisory Board.
Put it to work on your own notes

Try BastionGPT free for 7 days. HIPAA-compliant with a signed BAA on every plan, and your data is never used to train models.

Start your free trialRead the security overview

Updated July 2026