01Intake
Probabilistic where uncertainty lives.
Free-text notes, voice dictation, structured forms — three modes feed one extractor. Confidence is reported per fact, not buried. The model knows when it is guessing.
Verum · in private evaluationClinical inference engine · v0.1
Probabilistic
intake.
Deterministic
verdict.
Note
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 for payer-side workflows. LLMs read the clinical note. Pure code makes the decision. When the case has attached imaging, BRIDGE lets the reviewer have Verum · Visionread the image too, attest the findings, and promote them into the deterministic gate's fact bank — flipping verdicts on cases where the dictated report hedged on documentation. Every approval, pend, and denial carries a verbatim policy citation and a reversible audit trail — the way a medical director actually defends a case.
00The pipeline at a glance
- 01NOTEClinical noteFree-text, structured, or voiceProbabilistic LLM extraction
- 02FACTSExtracted facts14 typed fact spans
- 03GATEDeterministic evalRules × facts → verdict
- 04VERDICTDecision + auditAPPROVE · PEND · DENY
NOW INSPECTING — § 01
Probabilistic LLM extraction
01The instrument
Not a chatbot
wrapping a denial.
An engine that defends one.
Why this shape
The most expensive failure mode in a payer workflow is a non-deterministic verdict. The second most expensive is a verdict you can't defend. Verum is designed to make both impossible.
02Verdict
Deterministic where money lives.
Once facts are extracted, the verdict is pure code. Same facts, same policy, same answer — every run. No temperature. No regeneration. No surprise downstream of audit.
03Audit
Reversible where stakes live.
Every approval, pend, or denial is six components — rule id, criterion, verbatim quote, status, confidence, operator hand. Override one component, the verdict re-derives.
02bOne case
From intake to verdict.
A single prior-authorization case, fully evaluated. Note, extracted facts, deterministic verdict, and the citation register that defends it — all on one sheet.
Verum · 4231
verum · 4231
Case file · prior authorization
VERUM-4231
filed
2026·05·12
Member
Synthia Marlow
DOB 1962·03·15
MBR 8194·4231
Procedure · Payer
Lumbar MRI w/o
CPT 72148
Medicare Part B
Clinical note · excerpt
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.
Extracted facts · 3 of 14
- conservativeTx8.0w≥ 6w
- odi42≥ 30
- redFlagabsentbypass · n/a
Differential · imaging rationale
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.
Verdict · deterministic
APPROVE
confidence
0.94
Operator attestation
Reviewed by medical director on file. Reversible upon receipt of additional records. Standard appeal window: 60 days · Medicare Part B.
Citation register · click to view source
↗ 3 sources
··
One case. One verdict. Full audit, every time.
01bThree intakes, one engine
How does a case
arrive?
Free-text, voice dictation, or a 23-field structured form. One probabilistic extractor consumes all three. The clinical facts shape is the same downstream.
IDLE · click microphone to begin
Engine Groq WhisperP50 410ms
Raw audio · 16 kHz mono
noise floor · −42 dBDenoised · speech-only
✓ HP filter @ 80 Hz · spectral gateAuto-cycling demo. In the workspace, hold to record · release to transcribe.
Live transcript
Awaiting speech…
02At scale
Counts and
constraints.
The honest answer to "what is in the box." No vanity metrics, no rounded-up totals.
- — 010procedure familiesLumbar · Cervical · Brain · CT abdomen-pelvis (+ BRIDGE-extended lumbar decompression)
- — 020payer ontologiesMedicare · Commercial · Medicaid — two policy versions
- — 030LLM providers, one switchGroq · OpenRouter · Google · Mistral · NVIDIA
- — 040curated modelsReasoning-graded, with structured-output reliability
03Five moves
The pipeline produces
a decision. These produce action.
Five feature surfaces, each hooked into the same deterministic gate. Every output is signed by the same audit trail.
- BRIDGE
BRIDGE — image evidence into the gate
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.
3-8swarm-endpoint readOpen - Appeal
Appeal Generator
Denials draft their own rebuttal — paragraph by paragraph against the failing rules, signed by an operator, exportable as a structured PDF.
12statutes per letterOpen - Root cause
Root-Cause Analyzer
Every denial maps to a CARC/RARC code, a category, and a likely overturn estimate. Cohort dashboards distinguish systemic patterns from one-off denials.
47CARC codes cataloguedOpen - CDS Hooks
CDS Hooks Service
Spec-compliant FHIR endpoint at /cds-services — Epic, Cerner, and any conformant EHR see policy gates inside the order-sign workflow.
<50ms · cached invokeOpen - Eligibility
Eligibility Pre-Check
X12 270 / 271 ingest with coverage, cost-share, and prior-PA signal — before the order is placed, before the denial is written.
4destination bridgesOpen
04Provenance
Two ontologies are real.
The others are honest about it.
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.
4 procedures × 3 payers · 2026
LUMBAR · MCR
LUMBAR · COM
LUMBAR · MCD
CERVICAL · MCR
CERVICAL · COM
CERVICAL · MCD
BRAIN · MCR
BRAIN · COM
BRAIN · MCD
CT A/P · MCR
CT A/P · COM
CT A/P · MCD
Flagship REAL — verbatim from cited policy
Synthetic — flagged in sourceDocument
05Access
Open the workspace.
Run one case.
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.