Quality programs are only as serious as their handling of disagreement. This week we shipped end-to-end adjudication for the cases where independent raters do not agree, brought real-time grading analytics into the operational view, and removed a feature that lacked the clinical context to be safe.
What a serious quality program does with disagreement
Last week we shipped the infrastructure for blind grading. The honest measure of any quality program is what it does with disagreement, not how often it produces agreement. This week we closed that loop. When two independent raters disagree on a CTCAE grade, the platform now handles the adjudication end-to-end: detection, assignment, side-by-side review, e-signed resolution, and a written audit trail.
We also brought the operational metrics that previously required exporting data into the platform directly. And we made a deliberate product decision to remove a feature that we determined could not be shipped responsibly. Naming what we removed is as important as naming what we kept.
Here is what shipped, why it matters, and what is next.
Feature Highlight 1: Automated adjudication for grade conflicts
Grade disagreements between independent raters are inevitable in multi-rater workflows. The conventional handling is email, spreadsheets, and manual coordination to find someone qualified to review the conflict. That handling does not produce an audit trail. It produces a folder of forwarded emails.
The platform now detects the disagreement at submission time. If two raters submit different grades on the same case, the system assigns an adjudicator from the pre-designated pool. The adjudicator sees both grades, the clinical note, and the grade definitions in a single view. One click resolves the conflict under an electronic signature meeting 21 CFR Part 11 expectations.
Technical safeguards in the same release:
- Collision detection prevents duplicate adjudicator assignment during concurrent operations
- Load balancing distributes cases across the available adjudicator pool
- Digest notifications consolidate updates to reduce email noise
- Email sanitization to keep notification delivery reliable
The downstream effect: the FDA expectation of a documented resolution process for grading disagreements is met by the platform itself, not by a folder structure on someone's desktop.
Feature Highlight 2: Adjudicator role management
Before conflicts can be routed automatically, a pool of qualified adjudicators has to exist. The new role management surface lets coordinators designate team members as adjudicators through a dedicated sidebar, see current adjudicators with role badges and management controls, search and filter by name or email, and remove the role when team composition changes.
Authorization checks ensure only users with appropriate permissions can manage adjudicator assignments. The system validates the role at both the UI layer and the data layer.
Feature Highlight 3: Real-time grading analytics
The analytics dashboard now reflects live operational state across four tabs:
- Overview: active cases, grading volume, agreement, turnaround time
- Productivity: throughput, time-to-completion, coordinator efficiency
- Quality: inter-rater agreement, grade distribution
- Practice-wide: cross-organization metrics
The streams that feed this dashboard come from the same clinical workflows the platform runs, so the operational view is always current. Research directors can answer "what is our agreement rate this month?" and "how many cases are pending adjudication right now?" without exporting data or building external reports.
Feature Highlight 4: Encounter capture and CTCAE term search
Two interface changes that compound across a workday. Starting a new encounter from a patient profile now pre-populates patient context. The modal supports keyboard navigation for clinicians who never want to touch a mouse, and the state model makes the interaction snappy.
The CTCAE term search now spans the full set of over a thousand terms in a single searchable field grouped by system organ class. Selecting a term auto-populates the SOC. The previous two-step interaction (pick SOC, then find term) is gone.
A deliberate removal
This week we removed treatment recommendations from the multi-drug attribution workflow. This was a product decision, not a defect fix.
Clinical review identified that producing useful treatment suggestions requires patient context the platform does not have: comorbidities, contraindications, concurrent medications, patient preferences. A treatment suggestion produced without that context could be misleading. In the language we use to describe the engine: the constraint chain for treatment recommendations was incomplete, and shipping a feature that depends on a complete constraint chain is not a thing we do.
The principle stays the same: AI suggests grades and causality assessments. Treatment decisions remain with the clinician.
Knowing when not to generate output is a real architectural commitment. It is also one we are willing to enforce against features we have already built.
Improvements
- Manual grading entry for cases that do not need AI-assisted grading
- "Clinical Evidence" labels standardized to "Drug Interaction" across web and mobile
- Loading spinners replaced skeleton placeholders for more consistent state communication
- Completion checkmarks and collapsible causality details added to workflow tabs
- Patient list item colors improved for readability across light and dark modes
Bug fixes
- Patient list item color contrast now meets accessibility thresholds in both themes
- Review status updates correctly after grading submission
Looking Ahead
Next: adjudication-rate analytics, expanded agreement matrices segmented by SOC and rater pair, adjudication on the mobile provider app, and notification preferences for the digest frequency.
More Friday updates at burna.ai/blog.
