Radiation planning, orchestrated.
VerityRT unifies case intake, medical imaging, contouring, planning preparation, dose QA, evidence, documentation and treatment monitoring through specialised clinical AI agents — while preserving qualified human authority at the decisions that matter.
Every finding is explainable. Every source is traceable. Every calculation is reproducible. Every approval is recorded.
- Blocking
Oesophagus draft contour is discontinuous across four slices
An unintended gap of this size inside the high-dose region changes what is reported for this structure. The generating agent treated the gap as a genuine anatomical boundary; the independent recomputation does not support that.
Independent QA Agent · Structure Oesophagus · SYN-SS-0118-v1 agents disagree
- Review
Lung_L agreement with the prior approved version is below the protocol threshold
Most of the difference is at the inferior boundary, where the current scan shows a different diaphragm position. The change is plausible but larger than this pathway permits without review.
Contouring Agent · Structure Lung_L · SYN-SS-0118-v1
- Clinical Orchestratorv0.4.2Blocked
- Intake Agentv0.6.0Passed
- Imaging and 3D Agentv0.7.1Needs review
- Evidence Research Agentv0.5.3Needs review
- Contouring Agentv0.9.4Needs review
The live application, running on a synthetic case. Of 163 automated checks, two need a person — and one of them is a disagreement between the agent that drew the contour and the agent that checked it.
One coordinated intelligence layer, not fifteen chatbots
A radiation-oncology department already has the data it needs. What it does not have is something that assembles the case, checks the work against its own protocols, explains every finding, and produces the paperwork — reliably, on every case, without being asked.
Specialised, not general
Fifteen services with narrow purposes, explicit tool allowlists and a ceiling on what each may do. Dose and geometry come from deterministic code. Retrieval carries claim-level citations. A central orchestrator routes work and enforces policy; it does not decide medicine.
Exception-based, not step-by-step
A case that is going well should not need a person. Routine collection, reconciliation, routing, deterministic checks and documentation happen in the background. What reaches a clinician is the disagreement, the missing input, the anomaly and the decision.
Checked by something that disagrees
The Independent QA Agent uses separate rules, separate code and separate models from the agents that generated the work. It can block. It cannot approve anything, and it cannot dismiss its own findings. Where the two disagree, both positions are shown rather than averaged.
Referral to monitored treatment
Five human gates across the entire workflow. Everything between them runs on its own.
- 01ReferralRecords, imaging and documents arrive. The case is assembled and reconciled automatically.
- 02Case readinesshuman gateThe first of five human gates. A physician confirms identity, diagnosis, stage and that the workup is sufficient.
- 03Simulation and imagingImage quality control, registration confidence and a 3D representation, with no user input.
- 04Prescriptionhuman gateThe directive, the protocol and the record are reconciled field by field. A physician signs.
- 05ContouringDraft organs at risk with spatial uncertainty, compared against prior approved versions.
- 06Contour approvalhuman gateA physician approves the anatomy that everything downstream depends on.
- 07Planning preparationObjectives, geometry and variants, each with the tradeoff stated in plain terms.
- 08Dose and QADeterministic metrics, recomputed by a second independent implementation.
- 09Plan and physics QAhuman gateTwo independent physicist signatures. Unresolved critical findings refuse the signature.
- 10Treatment readinesshuman gateAuthorisation is handed to the existing licensed treatment workflow.
- 11Monitored treatmentPlan identity, setup and deviations are checked at every fraction. The monitor can recommend a hold. It has no beam control.
Fifteen specialised services
Each has an owner, a version, a tool allowlist, a set of stop conditions and a maximum autonomy level it can never exceed.
A0 observes. A1 does reversible administrative work. A2 recommends and a person decides. A3 prepares a change in a controlled workspace. A4 releases clinical state and A5 affects delivery — neither is granted to any service in this platform. Full agent catalogue
Modules
Each module is a separate scope with its own evidence and its own regulatory position.
- Plan QA
- Deterministic plan and chart checks against your own versioned protocols, with a second independent implementation for critical metrics. Plan QA
- Contour QA
- Geometric checks, spatial uncertainty, comparison against prior approved structures, and structured capture of every expert correction. Contour QA
- Evidence
- Claim-level citations with version, review date, jurisdiction, strength and local adoption. Conflicting sources are shown, not blended. Evidence
- Submissions
- Clinical review packets and prior-authorisation packages assembled from the case record, with the missing items named. Submissions
- Delivery monitoring
- Approved-plan identity, setup and deviation monitoring across a course, with a safe hold recommendation and no beam control. Delivery monitoring
What we will not claim
The fastest way to lose a physicist is to overstate what the software does. So here is the boundary, stated first rather than buried.
VerityRT does not
- Diagnose disease, or present an unconfirmed imaging finding as a diagnosis.
- Prescribe radiation or change a prescription.
- Approve contours, plans or physics review.
- Calculate a clinically authoritative dose.
- Write to a treatment-planning system or a record-and-verify system.
- Enable, control or authorise a beam.
VerityRT does
- Assemble and reconcile the case, and name what is missing.
- Run your approved deterministic checks on every case, the same way each time.
- Retrieve evidence with claim-level citations and show conflicts.
- Draft contours and planning objectives for review in a controlled workspace.
- Recompute critical numbers with a separate implementation and show any disagreement.
- Draft the documentation and record every decision immutably.
VerityRT is not FDA-cleared and makes no claim of clinical validation. Its regulatory status depends on the intended use finally adopted, and that assessment is ongoing. Read the clinical-safety position
Standards in, nothing out
VerityRT is standalone. It does not script a planning system or read a vendor database. A case arrives as standard radiotherapy objects through one interface, and there is no write path — not a flag that is off, a method that does not exist.
- Imaging and RT objects
- DICOM and DICOMweb as the interchange baseline: image series, structure sets, plans, dose grids, histograms and registration objects.
- One ingest boundary
- Five methods on a single interface. A site implements one file, wrapping an established DICOM library rather than a parser written here.
- Integrity before interpretation
- Frame of reference, plan-to-structure-set, dose-to-plan and histogram-to-structure references are checked before any agent sees the case. A blocking problem halts it.
- Clinical and administrative
- FHIR and HL7 v2 where appropriate, mapped into a canonical internal model that belongs to no single vendor.
- Nomenclature and notation
- TG-263 structure names and ICRU 83 dose notation, parsed rather than looked up, so a clinic can state a constraint the system has never seen without a code change.
No vendor source code, algorithm, documentation, screen layout or trade dress is reproduced anywhere in this platform. How interoperability works
What a department would measure
Illustrative figures from the synthetic cohort shipped with the prototype. They show which numbers matter, not what your clinic would see.
- 38 h
- median planning turnaround, down from 61 h across nine synthetic weeks
- 94%
- of cases reached the next gate with no manual administrative step
- 1 in 163
- checks on a typical case needed a person
- 100%
- of state changes carried an audit event
These are figures from a demonstration dataset built for this prototype. They are not measured clinical results, not a performance claim, and not evidence of benefit.
The honest way to evaluate this is to try to break it.
Open the synthetic demo and go straight to the case with the conflicting prescription. Try to approve it. Try to clear the critical finding as the wrong role. Try to bundle it away. The refusals are the product.