A grade-only blind grading mode, and a blind the system enforces

Week 26 of 2026 (22 to 28 June) added grade-only blind grading: the grading workflow proposes the CTCAE term for each adverse event, and each human rater assigns only the severity grade, without seeing the grade the workflow proposed. The blind itself moved into the query layer, so a rater is shown only the events they have graded. At the close of the week, on 27 and 28 June, organization admins gained an audit log explorer and an in-product compliance control register.

The work serves trial teams running multi-rater grading, and the admins who answer a quality or security review.

What is grade-only blind grading?

Blind grading means two or more independent raters grade the same adverse event without seeing each other's assessments. Every case used to run one way, with each rater choosing the CTCAE term and assigning its grade. Grade-only is a second mode.

Question	Term and grade (the default)	Grade only
Who chooses the CTCAE term	Each rater	The grading workflow, extracted from the note
Who assigns the grade	Each rater	Each rater, without seeing the workflow's grade
What the rater judges	The term and its severity	Severity alone: is this grade 2 or grade 3

Because raters stay blind to the proposed grade, their grades stay independent of the model and of each other, which is what an inter-rater study measures. How agreement is measured explains the method behind that kind of study.

Every event the workflow generates carries its provenance: the model, the workflow run, the timestamp and the electronic signature. A run retries automatically a bounded number of times, and admins can retry a failed run by hand. A grade outside the CTCAE range of 1 to 5 is dropped before saving, so one malformed value cannot sink a batch of good ones.

How a trial chooses its grading mode

Admins set a grading mode for the organization, and a single case can override it. The order is fixed: the case's own setting first, then the organization default, then term and grade. That rule lives in one module, so every screen that asks how a case was graded gets the same answer.

Blind grading enforced in the query, not on the screen

A blind-grading case holds events from both raters. A rater's case list now returns an adverse event only after that rater has submitted a grade for it, so neither rater sees the other's chosen terms before adjudication. The rule lives where the data is read, and an end-to-end test checks that each rater sees only their own work. Manually added events now take grades 1 to 5 only.

Two related changes landed alongside it:

• One Finalized filter. Admins see consensus and adjudicated cases together under a single Finalized filter. Finalized is a filter and never a stored status, so a case cannot drop out of every list.
• Patient records deny by default. Access to a patient's comorbidity data is refused when the patient has no organization or belongs to a different one.

An audit log explorer for organization admins

On 27 June, organization admins and owners gained an audit log explorer over every recorded action in their organization. They filter by action, status, role and resource, read volume on a time-series chart, and open any single event in a detail view.

The filters live in the page address, so one user's actions on one resource over one week becomes a link to bookmark or send to a colleague. Platform administrators gained the same view over system logs.

A compliance control register in the product

On 28 June, the compliance control register arrived in the product. A build pipeline reads five official sources and joins them into a register of 40 controls, each mapped across the frameworks:

• the NIST SP 800-53 moderate baseline
• the HIPAA Security Rule
• NIST SP 800-66
• the NIST Cybersecurity Framework 2.0
• the SOC 2 Trust Services Criteria

Each control carries a coverage assessment (covered, partial or not covered) and the evidence files behind it, and a test confirms that every cited file exists. In the admin area, the register groups controls by NIST family. Organization admins and owners set each control's status, and every change is written to the audit record.

The mapping to the SOC 2 Trust Services Criteria is a reference for a security review. Our SOC 2 Type 2 audit is in progress. The architecture a security review asks about is set out on one page.

Self-hosted model routing and steadier voice capture

Grading inference can now route to a self-hosted model when that route is switched on, with the hosted provider as the fallback when it is off. Every request to the self-hosted model pins the exact model revision, so the model behind a grade is known down to the revision.

Voice capture during patient interviews keeps the microphone stream alive across pause and resume, so the playback preview holds every segment recorded after a pause. When the microphone is unavailable, the clinician sees a clear message.

Since then: The following week, the on/off switch for the self-hosted route became a configured default model. In July 2026 the web provider app was retired; blind-grading administration, the audit log explorer and the compliance control register now run in app.burna.ai.

AI suggests, clinicians decide. Human-in-the-loop, always.

To see grade-only blind grading or the control register set up for your organization, book a 15-minute conversation.