Two layers: an LLM that reads, a rule engine that decides. Each is best at what the other is bad at.
Verum is a clinical inference engine. LLMs read the unstructured input — clinical notes, voice dictation, structured forms, attached imaging. Pure code makes every load-bearing decision underneath. The same gate runs payer-side prior-auth in Decide, provider-side multimodal CDS in Insight, and the relational data plane both pull from in Record. When a case has imaging, Verum · Visionreads it and the findings promote into the gate's fact bank — text-only and image-only claims arbitrate against the policy alongside each other. Every output carries a verbatim citation and a reversible audit trail.
Three product surfaces. One deterministic gate underneath. One custom-built clinical vision-language model (Verum · Vision) shared by both consumers.
3-agent pipeline (extractor → policy-match → deterministic gate) over per-payer / per-policy-version LCD rule trees. Trinary verdict with verbatim citations, calibrated confidence, counterfactual analysis. BRIDGE workflow inside /review promotes image findings into the policy gate.
6-stage pipeline (extract → annotate → image → fuse → gate → synthesis) over 4 flagship clinical-decision-support policies (CXR critical findings, ICDR retinopathy, NICE NG12 melanoma, Gleason ISUP). Verum · Vision is the optional image-extract layer; text path completes without it.
Two surfaces — Ingest (modality-side capture, parse, de-identify, render, content-hashed cloud persist) + Vault (relational record of studies, series, instances, scan observations, ICD-10 findings, audit log, dashboard, search, CSV export). The substrate Decide + Insight pull from.
The shape below is the canonical pattern. Decide instantiates it as a 3-agent prior-auth pipeline; Insight as a 6-stage multimodal pipeline with an image-extract layer; Record as a deterministic ingest-to-archive path with the same audit-everything contract. One pattern, three instantiations.
Each product is the pattern instantiated against a different question. Same extractor family. Same audit ledger. Same deterministic gate. Different verdicts.
Three-agent pipeline over per-payer LCD rule trees. Trinary verdict (approve / pend / deny) returned with the exact policy sentence the rule cites — not paraphrased, not summarized.
Six-stage pipeline runs the note through the text extractor and the image through Verum · Vision in parallel. The fuse stage labels each fact BOTH-AGREE / TEXT-ONLY / IMAGE-ONLY before the gate sees it.
Relational catalog of patients · studies · series · instances · scan observations · ICD-10 findings · audit log. FK + composite-PK enforced at the database level. Cloud storage paths are content-hashed, so re-uploads are idempotent.
The most expensive failure mode in any clinical workflow is a non-deterministic decision. The second most expensive is one you can't defend. Verum is designed — across all three product surfaces — to make both impossible.
Free-text notes, voice dictation, structured forms, attached imaging, DICOM uploads — five intake modes feed one extractor family. Confidence is reported per fact, not buried. The model knows when it is guessing.
Once facts are extracted, the gate is pure code. Decide arbitrates a prior-auth verdict; Insight arbitrates a clinical-decision recommendation; Record arbitrates referential integrity. Same facts, same policy, same answer — every run. No temperature. No regeneration.
Every output — a verdict, a finding, a soft-deleted scan — is six components: rule or constraint id, criterion, verbatim quote, status, confidence, operator hand. Override one component, the output re-derives. Nothing decided here is opaque.
A single prior-authorization case, fully evaluated. Note, extracted facts, deterministic verdict, and the citation register that defends it — all on one sheet.
62F presents with chronic low back pain × 9 weeks. Completed 8 weeks PT with structured pain program. ODI score 42. Failed NSAIDs and muscle relaxants. No red-flag findings on neurological exam. No bowel/bladder dysfunction, no saddle anesthesia. Provider requests MRI for surgical planning.
Disc herniation (L4-L5, L5-S1) vs. facet arthropathy vs. spinal stenosis. MRI required to establish anatomic correlate to the clinical findings before surgical referral.
Reviewed by medical director on file. Reversible upon receipt of additional records. Standard appeal window: 60 days · Medicare Part B.
One case. One verdict. Full audit, every time.
Free-text, voice dictation, or a 23-field structured form. One probabilistic extractor consumes all three. The clinical facts shape is the same downstream.
The honest answer to "what is in the box." No vanity metrics, no rounded-up totals.
Five feature surfaces, each hooked into the same deterministic gate. Every output is signed by the same audit trail.
The dictated report hedged. Verum · Vision reads the image. The reviewer attests + promotes the findings. The deterministic gate re-derives. Appeal letter + SOAP note auto-cite the evidence with reviewer attribution.
Denials draft their own rebuttal — paragraph by paragraph against the failing rules, signed by an operator, exportable as a structured PDF.
Every denial maps to a CARC/RARC code, a category, and a likely overturn estimate. Cohort dashboards distinguish systemic patterns from one-off denials.
Spec-compliant FHIR endpoint at /cds-services — Epic, Cerner, and any conformant EHR see policy gates inside the order-sign workflow.
X12 270 / 271 ingest with coverage, cost-share, and prior-PA signal — before the order is placed, before the denial is written.
Decide's pipeline is text-only because PA workflows arrive as a referral packet of notes. Insight runs in clinic at the moment of decision — the imaging study is on the screen already. Adding image-extract + a fuse stage lets the policy gate see both sides of the same patient.
CXR critical findings against the ACR Practice Parameter. Diabetic retinopathy + DME against the AAO ICDR 2003 4-2-1 grading. Cutaneous melanoma 2WW against NICE NG12 with the Glasgow 7-point checklist. Prostate Gleason scoring against ISUP 2014 v2 with documented 2014/2019 divergence.
The note runs through the text extractor. The image runs through Verum · Vision. The fuse stage labels every fact BOTH-AGREE / TEXT-ONLY / IMAGE-ONLY before the deterministic gate sees it. The gate uses the same evaluator (evaluator-fused.ts) the Decide gate uses — same rule schema, same severity-tier logic.
Every accept / promote / override / revert / resolve-dispute action lands in an append-only InsightAuditRecord with sealed-state immutability (SealedRecordError) and supersedes-chain lineage. Sealed records export to react-pdf and to a FHIR R4 Bundle containing Patient + Condition[] + DocumentReference + Provenance + AuditEvent[].
The other two products run on JSON files and sessionStorage — appropriate for their workflows. Record holds the longitudinal study graph, the audit history, and the bytes themselves — that needs FK enforcement, composite-PK constraints, and a real query language underneath.
patients · studies · series · instances · scan_notes (append-only) · instance_findings (composite PK on sop_uid + icd10) · audit_log · 4 reference tables. The 4-step hierarchy enforces referential integrity at the database level. Deletes require a typed reason; restores re-link cascaded children.
48-tag DicomTagSet schema. Deterministic FNV-1a 64-bit hashing on the 5 PHI tags PS3.15 Annex E lists (PatientName / PatientID / PatientBirthDate / IssuerOfPatientID / OtherPatientIDs). WL transform + paintToCanvas with MONOCHROME1 inversion shared between Ingest, Vault hover-preview, and Vision Chat DICOM upload.
Global search across patient / study / series / instance / scan observation / ICD-10 finding — the finding search hits code + description + synonyms in one query. Dashboard panels: modality donut + 30-day storage sparkline + 84-day activity heatmap, all pure SVG, no chart library. Hover-preview thumbnails on every row use a module-scope LRU cache to keep render under 100ms.
Lumbar and cervical Medicare 2026 are encoded verbatim from Aetna CPB 0236, cross-validated against CMS LCD L34220 and Carelon Imaging of the Spine. The other ten ontology cells are synthetic — every fact carries an explicit SYN- prefix.
No hidden mocks. No "AI-generated" hand-waving.
Authorized accounts open into the evaluation workspace immediately. Teams without credentials reach the engine through a brief alignment call — Verum is in private evaluation, by design.