Responsible AI · Architectural Commitments

The architecture is the responsibility.

Burna AI's commitments to responsible AI in oncology safety are not policy statements layered on a generative model. They are properties of the twelve-agent cascading constraint pipeline. Citation-bound output. Human-in-the-loop at every signed record. Named algorithms with visible reasoning. No model training on customer data without explicit consent. Two patents filed on the architecture.

Read the engine architecture Download the Architecture and Implementation Brief

PostureArchitectural, not policy

Review cadenceQuarterly · plus on architecture change

Last reviewedMay 2026

Citation-bound by design. Human-in-the-loop on every signed record.

01 / Five architectural commitments

Load-bearing properties, not policies.

Each commitment is a property of how Burna AI is built. None of them is enforced at runtime. None of them is a guardrail layered on a generative model. They are properties of the cascading constraint pipeline.

Citation-bound output

The engine cannot return a grade without a source sentence in the clinical note and a matching CTCAE criterion in the retrieval index. This is enforced at the architectural layer, not at the prompt or the output filter. There is no path through the pipeline that produces a grade without citation.

Hallucinations in clinical AI are usually presented as a tradeoff between model capability and safety guardrails. Burna takes the tradeoff off the table. The engine cannot hallucinate a grade because nothing in the architecture allows a grade to exit without citation.

Human-in-the-loop on every signed record

The engine produces a suggested grade with its citation, its evidence chain, its named-algorithm attribution, and its confidence calibration. The clinician reviews. The clinician modifies, approves, or rejects. Only the approved and signed version becomes the record. The 21 CFR Part 11 e-signature trail is captured on every modification.

The engine never signs. The engine never finalizes. The engine never acts autonomously on clinical data. This is the architectural answer to "AI suggests, clinicians decide."

Named algorithms with visible reasoning

Causal attribution runs WHO-UMC and Kramer causality algorithms against the evidence record. The output is a ranked attribution with the algorithmic reasoning path visible to the clinician, monitor, sponsor, and regulator. The output is not a confidence number from a generative model; it is a named algorithm with a defensible reasoning chain.

Visible reasoning is the architectural answer to "the black box."

No training on customer data without explicit consent

Customer clinical text is processed for inference only. The grading engine does not train, fine-tune, or learn from customer data unless that customer has executed an explicit written consent for that specific use, under explicit governance review, with clear scope and revocation rights.

The default is no training. The default cannot be changed by a vendor policy update; it requires an executed agreement.

PHI handling is architecturally constrained

Patient identifiers stay inside the customer's network in on-premises deployments and inside the customer's jurisdiction in regional cloud deployments. The grading model only sees de-identified clinical text. Zero PHI in logs is enforced by pre-commit checks in the engineering pipeline, not by post-hoc audit.

This is the architectural answer to "where does the patient data go."

02 / Eight exclusions

What the engine cannot do.

Commitments by exclusion. Each names a thing the engine architecturally cannot do, not a thing it has been instructed not to do. The architecture cannot do these things.

The engine cannot Produce a grade outside the defined CTCAE set.

The engine cannot Skip citation.

The engine cannot Contradict its own upstream findings.

The engine cannot Finalize a record without clinician signature.

The engine cannot Train on customer data without written consent.

The engine cannot Share data with third parties for advertising or remarketing purposes.

The engine cannot Operate in a fully autonomous mode that removes clinician review.

The engine cannot Be deployed in a configuration that bypasses the audit trail.

Each exclusion is an architectural property, not a policy commitment.

03 / Governance and review

Internal engineering, external oversight, customer governance.

Three nested governance loops. Each operates independently. Each is documented. Each is auditable.

Internal

Engineering review for every architecture change that touches the constraint pipeline. Clinical advisor review for every change affecting grading suggestion logic. Pre-commit checks for PHI exposure, hardcoded credentials, and disabled tests. Audit log preservation for every grading decision, modification, and signature.

External

Independent Review Board (IRB) or equivalent oversight for every validation study. Peer-reviewed publication path for every accuracy claim. Conservative reporting in all external materials. 21 CFR Part 11 alignment for the electronic record system. HIPAA Business Associate Agreement with every customer handling PHI. GDPR and regional privacy framework alignment in international deployments. SOC 2 Type II audit in progress (Q2 2026 kickoff). FDA EDSTP application in review.

Customer

Each design partner pilot is co-designed with the customer's IRB, ethics committee, and clinical operations leadership. The 12-pillar Implementation Playbook includes governance and policy guidance for the Office of Research and CISO. Site-specific calibration and tuning happens only under explicit governance review at the customer institution.

04 / Bias, fairness, equity

Where bias can enter, and where the architecture constrains it.

Adverse event grading involves clinical judgment about severity, attribution, and causality. Bias can enter through training data, evaluation cohorts, the criteria themselves, or how clinicians at different sites apply the criteria.

The architecture’s posture

Citation-bound: every grade is anchored to a specific source sentence and a specific CTCAE criterion. A reviewer can verify the grade against the source text directly. Bias hidden in a generative model's reasoning cannot survive the citation requirement.
Named algorithms: WHO-UMC and Kramer attribution are peer-reviewed, published algorithms. Bias in attribution is constrained to bias in the published algorithm, not bias in an opaque model.
Multi-site visibility: real-time grading consistency across sites surfaces site-specific patterns. If one site systematically grades differently from peers, the architecture makes that visible while the trial is running.
Human-in-the-loop: every signed record passes through a human clinician. Bias in the suggestion can be corrected at the point of signature.

What we are still working on

Validation cohorts at sufficient demographic and geographic diversity to characterize performance across patient populations.
Peer-reviewed publication of fairness analyses across patient subgroups.
Ongoing review by the clinical advisor bench for sites of bias the architecture does not yet address.

This statement will be updated as validation data publishes.

05 / Transparency

Publicly available, and what remains proprietary.

The architecture, the validation roadmap, the compliance framework, and this statement are publicly available. Specific implementation details, retrieval index construction, internal calibration thresholds, and accuracy percentages until peer-reviewed publication remain proprietary.

Publicly available

The engine architecture, with the twelve agents named and constraint relationships described.
The validation roadmap, with studies in flight, partners, and target publication venues.
The 5-page Architecture and Implementation Brief, available at security@burna.ai or in the Resources library.
The Query Economics Whitepaper (Pronker, Tolmie, Oracle ClearTrial citations).
The 12-pillar Implementation Playbook.
The compliance framework (HIPAA, SOC 2, 21 CFR Part 11, GDPR).
The subprocessor list.
This Responsible AI Statement.

Remains proprietary

The implementation of each agent in the cascading constraint pipeline.
The retrieval index construction.
The internal calibration thresholds for the confidence calibration agent.
Specific accuracy percentages and Kappa scores, until they publish in peer-reviewed venues.

06 / Feedback and disclosure

If you find a vulnerability, a bias, or a safety issue, we want to hear from you.

Three contact rails. Each routes to the responsible owner. Each commits to a defined response window.

Security vulnerabilities

security@burna.ai

Acknowledgement within 2 business days. Substantive response within 30 days where feasible. See the responsible disclosure policy for full scope and safe-harbor terms.

Responsible disclosure policy →

AI governance issues

responsible-ai@burna.ai

For bias, fairness, or AI governance concerns specifically. Same acknowledgement and response cadence as security reports.

Email responsible-ai@burna.ai →

Regulatory inquiries

partnerships@burna.ai

For FDA, EMA, MHRA, PMDA, NMPA, or equivalent regulatory authority inquiries. Include relevant authority and inquiry scope.

Email partnerships@burna.ai →

Last reviewed

This statement is reviewed at least quarterly and after any architectural change that materially affects how Burna handles clinical data, makes grading decisions, or governs validation. The current version is dated May 2026. Material changes will be notified via the quarterly briefing and via prominent notice on this page.

The architecture is the responsibility.

PostureArchitectural, not policy

Review cadenceQuarterly · plus on architecture change

Last reviewedMay 2026