A source-linked review after the AI draft

Compare six kinds of session substance with an AI-generated therapy progress note—then inspect the source and decide what belongs in the record.

A second pass that starts from the session

Note drafting asks, “How can this session become a useful note?” The omission check asks a different question: “What concrete material in the session might a note need to carry, and is that substance already present?”

The two passes are separate on purpose. The first stage of the omission check sees the transcript and does not see the note. It builds a bounded checklist from the source before the existing draft can anchor its attention. Only after each proposed excerpt is mechanically matched to the transcript does a second stage compare that checklist with the saved note.

Possible gaps return as a small review queue. Intuita does not insert them into the clinical record. “Added”, “reviewed” and “dismissed” are workflow decisions; the therapist remains the author and must make any warranted change in the note itself.

The checker is for individual sessions with a saved, non-empty note and a usable transcript. Couple sessions do not use this path.

Six categories, not an open-ended critique

The transcript-only pass may identify up to ten concrete items across six schema-bounded categories:

1. Safety-related content: explicit patient words about suicide, self-harm, abuse or violence.
2. Medication: a named medication, dose, adherence issue, start, stop or change.
3. Plan change: an explicit change in goals, frequency, referral or treatment plan.
4. Between-session work: an assignment, exercise or task.
5. Intervention: a specific therapeutic technique used in the session.
6. Key disclosure: a major factual disclosure not covered by the other categories.

The categories are designed around concrete documentation substance. They do not ask the model to grade the therapist’s writing style, diagnose the patient or turn every emotional statement into a missing-note warning.

Source before suggestion

Every proposed excerpt must be matched to the raw transcript before the existing note is examined.

The verifier normalises Unicode, case, Dutch diacritics, apostrophe variants, whitespace, speaker furniture and punctuation. It can search across one to three nearby same-speaker segments while tolerating a short backchannel from the other speaker. Safety-flavoured material receives a wider retry window.

If no source match is found, the candidate does not reach the note-comparison stage. The matched source segment supplies the speaker role and transcript pointer; those are not guessed from the proposed text.

“Source-matched transcript excerpt” is more precise than “byte-for-byte verbatim quote”. The candidate wording is retained after a tolerant match, so harmless differences in case, apostrophes, diacritics, spacing or punctuation can remain. A paraphrase does not pass this source gate.

The category and short explanation are AI-generated interpretation. The excerpt is AI-selected and mechanically matched. Those are different kinds of provenance.

Meaning, not exact-word matching

After source verification, a second comparison receives only:

• the verified checklist;
• the flat text of the saved note.

It asks whether the note already contains the same substance. Exact wording is not required. If the note paraphrases the material adequately, the item counts as covered and disappears from the possible-gap queue.

There is also a narrow mechanical number check. For medication or other digit-bearing material, an item can be treated as covered when every digit from the excerpt appears somewhere in the note. This is a cost-saving shortcut, not semantic proof; the final result still requires therapist review.

The checker is content-oriented, not format-oriented. It does not know whether every SOAP, DAP, BIRP or other section is structurally complete. It asks whether selected session substance appears somewhere in the note text.

Precision before volume

Covered items are removed. Low-confidence missing judgments are removed. The remaining non-safety items are prioritised in this order:

1. medication;
2. plan change;
3. between-session work;
4. intervention;
5. key disclosure.

Normally no more than five total flags are shown after that prioritisation. Verified medium- and high-confidence safety-category items are exempt from the final five-item cap.

This design prefers a small inspectable queue over a long speculative critique. It also means “no flags” cannot prove completeness. It means only that nothing survived extraction, source matching, containment comparison and confidence gates.

Safety without a risk score

Safety flags are based on explicit source language. The checker does not infer risk from affect, topic or diagnosis and does not assess intent, immediacy or severity.

When a verified medium- or high-confidence safety-category item survives:

• it is not removed by the final five-item cap;
• it cannot be dismissed like an ordinary flag;
• the therapist can mark it added or complete an explicit reviewed-against-session confirmation.

Low-confidence candidates are removed in every category, including safety. An excerpt that cannot be matched even after the wider safety retry is also withheld and escalated internally. The accurate promise is not “safety signals are never filtered”. It is: verified medium- and high-confidence safety flags receive protected review handling.

That handling does not replace clinical risk assessment or the therapist’s safeguarding obligations.

What the therapist sees and controls

Each visible card can contain:

• one of the six categories;
• a source-matched transcript excerpt;
• the matched speaker role and segment;
• an AI-authored statement of what may be absent;
• medium or high confidence;
• a stable identity used to preserve review state.

Cards open with their evidence visible. A therapist can:

• mark an item added after editing the note;
• dismiss an ordinary item;
• mark a safety item explicitly reviewed.

“Added” does not copy AI text into the note. It records the therapist’s workflow status. The note changes only when the therapist edits it.

A repeated flag keeps its status when category and normalised source excerpt produce the same stable identity. That supports re-checking without reopening every unchanged item.

Nine progress-note formats — one substance check

The live note editor supports nine AI-fillable progress-note formats: SOAP, DAP, BIRP, GIRP, PIRP, SIRP, PIE, Narrative and EMDR. Longitudinal document templates are separate.

The omission checker itself is format-agnostic. It compares a transcript-derived checklist with flat note content. It does not inspect a format definition or decide that every required heading is complete.

EMDR’s SUD and VOC safeguards belong to draft generation, not the omission checker. Those therapist-confirmed numeric fields are excluded from the model’s response schema, so AI drafting cannot populate them. The second pass does not independently validate those measurements.

Current format-scope limitation

Today, the note read used by Cortex is keyed to the session but not additionally to the requested note format. Because the backend can store more than one format for a session, a multi-format session can cause the check to compare the transcript with a different saved format while attaching the returned flags to the requested one.

Until that binding is corrected, the defensible use is:

• treat the result as a review aid, never proof;
• be especially cautious when more than one note format exists for the session;
• re-check after saving the intended note;
• inspect the excerpt and the actual note before accepting any flag.

This is a live limitation, not a theoretical edge case, and the public explanation should not hide it behind “checks your current note”.

Asynchronous and imperfect by design

The checker normally runs after a successful draft and can also be requested manually. A manual re-check flushes pending note changes before dispatch. The interface polls for the asynchronous result.

The result is not version-fenced to the exact note revision. If the note changes while a check is running, or two checks overlap, a late result can describe an older draft. Re-checking after edits is therefore important.

Other limitations include:

• transcription errors can affect both checklist and source matching;
• an AI extractor can miss relevant material;
• source verification can withhold a genuine item when wording does not match strongly enough;
• containment can mistake a mention for adequate coverage or miss a good paraphrase;
• a malformed internal result can sometimes collapse toward an empty queue;
• after the polling window, a delayed result may not immediately appear;
• generated explanations are prompted and checked to remain observational, but structural checks do not make them clinically correct.

The surface’s “No flags found — not a completeness guarantee” is the right reading.

What this capability can safely say

It can say:

• a separate AI pass started from the transcript rather than the note;
• each displayed excerpt was mechanically matched to session source language;
• a second pass judged that the same substance may not be present in the saved note;
• low-confidence and already-covered items were removed;
• the therapist decides whether and how to change the record.

It cannot say:

• the note is complete when no flag appears;
• every safety disclosure will always be shown;
• every excerpt is byte-for-byte raw transcript text;
• the correct format was certainly checked in a multi-format session;
• accepting a flag automatically updates the note;
• the checker understands every format requirement;
• an AI statement should be copied into the clinical record without therapist verification.

Synthetic evidence trace

Source-matched excerpt: “We agreed I would call the prescriber tomorrow.”
Generated possible omission: “The agreed follow-up may be absent from the note.”
The excerpt is AI-selected and mechanically source-matched; the omission statement is AI-generated. The therapist checks both the source and the actual saved note.

Evidence, implementation and review status

• Official audit contrast: CMS behavioral-health documentation self-audit.
• Research context: ambient-AI note quality study.
• Product implementation reviewed against the repository on 25 July 2026; this is not a compliance audit or completeness validation and no independent clinical reviewer is claimed.
• Current rendered limitations: the saved-note read is not safely bound to requested format or exact revision; late results can be stale, and malformed results can collapse toward an empty queue.

Review the therapy-note workflow or ask about the current implementation.