Adverse event provenance: AI-suggested or provider-added, and edited or not

Adverse event provenance records how each event entered the safety record: suggested by AI or added manually by a provider, and accepted as suggested or edited first. In week 24 of 2026 (8 to 14 June), Burna AI began keeping both facts on every graded event, scoped blind-grading rater and adjudicator roles to each organization, and locked an adjudication case to the first adjudicator who claims it. Two features were removed on purpose the same week: the ICD-10 panel in CTCAE search and the Rejected review status.

AI vs manual adverse event provenance on every graded event

Two facts now travel with every graded adverse event:

Field	What it records	When it is set
Source	Suggested by the grading workflow, or added manually by a provider	When the event enters the record
Edited	Whether a suggested grade was changed before acceptance	When a provider edits it, and kept after acceptance

An edited event shows as Edited while it waits for review and keeps an edited mark once accepted. The record reads "Grade 3, suggested by AI, edited, accepted" where it once read "Grade 3, accepted".

A grade and its attribution are the investigator's decisions, and a monitor, an auditor or a sponsor reviewing the record needs to see where the AI's suggestion ended and the provider's judgment began. AI suggests, clinicians decide, and the record now shows how each decision played out. The full list of what is captured is in the FAQ on what the Burna AI audit trail includes, and the controls behind it are on the architecture a security review asks about.

Organization-scoped grading roles for blind grading

Blind grading separates the raters who grade a case independently from the adjudicator who resolves their disagreements. This week the rater and adjudicator roles moved from the user to the organization membership.

• Roles per organization. A clinician can rate cases in one organization and work as a coordinator in another. An organization admin grants or revokes the adjudicator role within that organization; a platform admin can do so across organizations.
• Rater role derived from assignments. Assigning someone as Rater A or Rater B on a case makes them a rater in that organization, and clearing their last assigned case releases the role, so there is no separate switch to keep in sync.
• An adjudicator queue. Adjudicators see the cases awaiting them, with a count of the events where the raters disagree. Rater identities appear as Rater A and Rater B.
• Calibration case sets. Platform admins can load a set of 50 sample CTCAE cases into an organization for blind-grading practice and clear it afterwards. Cleared cases disappear from every list and queue, and their rows are kept for the audit trail.

Blind grading, adjudication and agreement measurement appear in the dated record of what Burna AI does today.

Adjudication: the first claim locks the case

• First claim wins. The first adjudicator to act on a case claims it, and a second adjudicator sees the case as locked.
• Conflict-of-interest guard. A person who rated a case cannot adjudicate it. The server enforces the rule, and the adjudication queue leaves those cases out.
• Case state moves on its own. When both raters submit, the case goes to consensus if their grades match and to pending adjudication if they differ.
• Submission timestamps. Each rater submission records when it was submitted and when it was last updated, shown as "Submitted on" and "Updated on".
• Deleting a grade before submission. A rater can remove their own grade on a single event until they submit. After submission the grade is locked, and other raters' grades on the same event are untouched. Each rater's progress moves from not started to in progress to submitted automatically.

Two deliberate removals: the ICD-10 panel and the Rejected status

• ICD-10 panel removed from CTCAE search. Providers found ICD-10 codes beside CTCAE terms confusing, and ICD-10 coding happens downstream in billing, separate from adverse event grading.
• Rejected review status removed. The review lifecycle went from three states to two, pending review and accepted, after clinicians described their practice: accept a suggested grade, or keep it pending for further review. Existing records marked Rejected read as pending review.

Data quality at the point of entry

• Date of birth in local time. Date-of-birth fields accept dates up to the user's own local date rather than today's date in UTC. Future dates and dates before 1900 are rejected in patient creation, encounter import, SAE reporting and EHR patient search, and a date picker with month and year menus replaces the browser's date field.
• Duplicate emails caught early. A patient's email is checked against the organization's existing patients and team members while the form is filled in, so a duplicate shows inline before submission.
• One source for demographics. Patient details read from the canonical patient record rather than a separate copy.
• Encounter filters on the server. The encounter list filters by date range, patient, status and keyword on the server, with a patient picker that searches as you type.

Voice capture from the first word

Audio capture now starts the moment a recording starts, rather than when speech is first detected, so the opening words of a patient's or clinician's recording are kept.

Also shipped in week 24

• Grading continues with partial context. When organization context is briefly missing, for example during a sign-in transition, grading proceeds with reduced context and logs it instead of stopping the run.
• Grade colors from one source. CTCAE grade colors are defined once and used on every screen.
• Seven languages for grading and dashboards. Interface text across the grading and dashboard screens is translated into English, Arabic, German, Spanish, French, Portuguese and Chinese.
• Study site from the trial. When a PRO enrollment token is created, the study site fills in from the selected trial.
• Uploaded documents stay with their owner. Organization document queries check the document's category and owner before returning a result, so each user sees only the knowledge documents they uploaded, and an organization admin no longer sees another member's private documents.

Since then: in July 2026 the web provider app was retired in favor of the platform at app.burna.ai, and since August 2026 a PRO enrollment token opens Healthnotes, the trial participant app.

AI suggests, clinicians decide. Human-in-the-loop, always. To see how provenance reads on a graded record, book a 15-minute conversation.