An adjudicator queue that fills from rater disagreement, and CTCAE v5.0 running beside v6.0

In week 35 of 2026 (24 to 30 August), the Burna AI adjudicator queue began filling from the raters' submitted grades, so a case reaches an adjudicator the moment two raters disagree, and organizations can now run blind grading adjudication end to end on their own. The full CTCAE v5.0 vocabulary went live beside v6.0, with a lookup that declines to substitute one version's term for the other's, and every 21 CFR Part 11 signing ceremony began resolving to a signature record. The work serves trial teams that measure agreement between raters, sites running protocols on both CTCAE versions, and the quality staff who answer for the signature and the audit trail.

How the adjudicator queue finds a rater disagreement

In a blind grading study, two raters grade the same case without seeing each other's work, and an adjudicator resolves the cases where they disagree. The disagreement is the measurement, so it has to reach the adjudicator. What blind grading is covers the method.

The queue used to filter on a case status that changed only after both raters ran an explicit finalize step. A case where two raters had submitted different grades could wait unseen because one of them had not finalized. The adjudicator queue now derives from the grades themselves:

• Membership comes from disagreement. A case appears when two raters submit different grades for the same CTCAE term, whether or not anyone has finalized.
• Conflicts of interest read the assignment. A rater is kept away from a case they graded by reading the assignment record, which exists whenever the assignment does.
• The scan is bounded. Cases are read in pages of 100 by cursor, so memory stays constant as the caseload grows and no case is skipped or counted twice at a page boundary.

Blind grading adjudication, run by the organization

Three changes let an organization run a study without help from platform staff. Granting adjudicator permission takes effect where it is checked. Raters, adjudicators, organization owners and administrators each reach their own view, scoped to their own cases. And a case advances the moment a rater submits:

What the raters have submitted	What happens to the case
Not every rater has submitted	In progress
Every rater agrees	Closed as consensus
Any disagreement on a term	Sent to adjudication

Each rater works one lifecycle, Draft, In progress, Submitted, over their own graded events and nobody else's. Submitting is an explicit, irreversible confirmation that needs at least one graded event, and the server enforces that rule as well as the screen.

Why measuring rater agreement matters

The published baseline shows how far apart careful reviewers can land. When two reviewers identified CTCAE toxicities in the same 100 radiation oncology notes, their agreement measured kappa 0.59 to 0.68 (Fairchild et al., International Journal of Radiation Oncology, Biology, Physics, 2020). In the placebo arms of two phase III trials, when the same adverse event was reported for the same patient at several visits, its attribution category changed at least once in 31 to 36% of cases (Hillman et al., Journal of Clinical Oncology, 2010). A blind read is how an organization measures that spread on its own cases, and how agreement is measured sets out the method.

CTCAE v5 and v6 side by side, without substitution

A trial keeps the CTCAE version its protocol names, so many open trials still grade against v5.0. This week the full v5.0 vocabulary went live beside v6.0:

	CTCAE v5.0	CTCAE v6.0
Terms	837	850
MedDRA release	20.1	28.0
System organ classes	26	26

Source: NCI CTEP, CTCAE v5.0 and v6.0.

Every v5.0 term carries an anchor to its page in the NCI document, so a grade can cite where its criterion is printed. The version is threaded through lookup, both grading workflows and the catalog search, and v5.0 became the default version.

Running two versions at once exposed a substitution risk. When a term exists in one version and not the other, a lookup that returns the nearest match from the other version keeps the workflow moving and puts the wrong term in the trial database. With v5.0 as the default, the abbreviation PLT had resolved to Thrombotic thrombocytopenic purpura. Three rules now govern the lookup:

1. Only genuine renames are bridged. Platelet count decreased in v5.0 is Thrombocytopenia in v6.0, and Lymphocyte count decreased is Lymphopenia.
2. A version gap is declined. A term that exists in the other version but not in this one is reported as a gap and not substituted.
3. A tie goes back to the clinician. Fuzzy matching ranks candidates by edit distance and answers only when the best match clearly beats the next one.

Why a CTCAE lookup returns ties instead of guessing

The CTCAE vocabulary is built from near-identical pairs. Scored against each other by edit distance, these different terms come out closer than a typical misspelling does to the term it was meant to be:

Near-identical pair	Similarity
Lymphocyte count decreased and Lymphocyte count increased	0.923
Skin hyperpigmentation and Skin hypopigmentation	0.909
Dysphagia and Dysphasia	0.889

No threshold separates a typo from a different term, so the lookup measures the gap to the runner-up. A misspelling such as "diarhea" resolves to Diarrhea. "dysphaia" scores the same against Dysphagia and Dysphasia, so both come back for a clinician to choose between. Criteria search also reads grade descriptors and definitions, so "ANC <1000" reaches Febrile neutropenia.

The trial-operations side of running both versions is in part 2 of the CTCAE v6.0 migration guide, every CTCAE v6.0 change, term by term is on ctcae.ai, and which CTCAE versions Burna AI supports answers the short question.

A Part 11 signature is a record at five decision points

Under 21 CFR 11.50, a signed electronic record shows the signer's printed name, the date and time of signing, and the meaning of the signature, and section 11.70 links the signature to its record so it cannot be transferred. This week:

• Every ceremony resolves a signature record. All four sign-off ceremonies resolve a signature record with its own purpose, rather than computing a local value.
• Five clinical decision points carry a signature. The pre-visit queue, a patient-reported outcome submission, grade review, grade adjustment and intervention recording each run the same signing ceremony.
• The audit entry is durable. The audit service writes each entry to durable storage first, and an entry that cannot be written raises an error instead of being dropped without a trace.

Attribution review, separate from grade review

Accepting a drug attribution used to accept the grade with it, because both decisions shared one review status. A clinician can now accept that a drug caused an event without accepting a grade they have not yet reviewed. Grade actions leave attribution alone unless the list of suspected drugs actually changed, and a clinician's attribution decision survives an AI re-run that produces the same list.

Encounter queues that count the whole queue

Encounter and grading queues used to filter one page of 100 rows in the browser, so a status chip counted only the rows that had loaded. Filtering, counting and paging moved to the server:

• Counts cover the queue. Each chip's count comes from the whole queue, not the page on screen.
• Pasted ids work. An encounter id pasted exactly as it appears on screen finds the encounter.
• Views are shareable. The filtered view lives in the page address, so a coordinator can send a colleague a link to exactly what they are looking at.
• Pages open at once. Titles, search fields and filters appear immediately while the rows load.

Also in week 35

• Workspace switcher. People who belong to more than one organization switch between workspaces from the account menu, and a new workspace's address is generated from its name.
• One enrollment vocabulary. Enrollment status is defined once and read the same way everywhere, so a patient's enrollment shows correctly to every clinician. A recording still in progress stays out of the provider queue until it completes.
• Dashboards that follow the record. Visit and patient-reported outcome analytics update after every clinical write, and grade counts come from the adverse events themselves.
• A second inference provider. Grading can run on an additional model provider, selected by configuration.

Since then: The following week, a trial began naming its own CTCAE version, every graded row began recording the version that produced it, and the version became a choice made before a grading run starts, as the week 36 update describes. ctcae.ai launched in September 2026 with every v5.0 and v6.0 term side by side.

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

To see blind grading adjudication, or both CTCAE versions running on your own protocol, book a 15-minute conversation.