The verdict is not a guess.
It is the projection of facts onto rules.
Pure LLMs cannot be a payer's medical director — the same case produces three different denials across three runs. Pure rule engines cannot either — they cannot read a free-text note. Verum extracts probabilistically and decides deterministically. Below: the gate, made visible.
Rule projection, live.
Three threshold rules. Three clinical facts. As you move each fact across its threshold, the verdict on the right re-derives — in pure code, the moment your finger lifts. This is how the real engine works. No model is called.
- conservativeTx ≥ 6w8.0w
- odi ≥ 3042
- redFlag bypassabsent
Three runs. Three verdicts.
Or one.
"62F with 8 weeks of PT-documented chronic low back pain. ODI 42. Conservative trial of NSAIDs and structured PT failed. No red-flag findings on exam. Requesting lumbar MRI without contrast."
"Conservative therapy possibly insufficient. Recommend additional records review."
"All mandatory criteria met. PT ≥ 6w (✓ 8w). ODI ≥ 30 (✓ 42). No red flags required for approval."
- · Rules evaluated in pure code
- · No temperature, no sampling
- · Same facts → same answer, always
A rule says ≥ 6 weeks. The fact says 5.5.
A rigid gate would deny. A clinically-aware gate routes to PEND@0.55 — "additional records needed" — when the fact is within ε of the threshold.
ε = max(2, threshold × 0.25) — calibrated so the borderline band scales with the threshold magnitude. Hard fails still dominate; only numeric mandatory rules participate.
Restores the legacy LLM "−0.15 borderline" tier, deterministically.
Every verdict is six things.
Override any one. It re-derives.
Rule identity
Every gate references a stable ruleId — e.g. LCD-L34220-3.1. The verdict ties to it; the rule's text can change without the verdict losing its anchor.
ruleId: stringCriterion
The rule's machine-readable threshold — ≥ 6 weeks, present/absent, ICD prefix match. Same shape across every payer ontology so cross-payer comparison works.
criterion: { kind, threshold, ... }Verbatim quote
The policy text the rule was encoded from — copy-paste from Aetna CPB 0236, CMS LCD L34220, or Carelon imaging. No paraphrase. Source of authority.
evidenceQuote: stringMet / unmet / borderline
The three states a criterion can produce. Borderline only applies to numeric rules within ε of threshold. Hard fails dominate aggregation.
status: "met" | "borderline" | "unmet"Confidence
A separate scalar from the fact extraction. Surface-able as a calibration metric in evals. Borderline verdicts cap at 0.55.
confidence: 0..1Operator hand
A medical director can promote any fact or rule outcome. The override is itself audited — original verdict + override + new verdict, all retained.
DecisionOverride · ResolutionAuditThe report hedged.
Verum reads the image.
Dictated radiology reports hedge. “L5-S1 narrowing on imaging, etiology unclear, recommend clinical correlation.” The gate pends on the documentation gap — surgical decompression authorization requires imaging-confirmed moderate-to-severe stenosis with documented etiology. Patient round-trips for a clarifying read; case takes another week.
BRIDGE closes that loop. Inside /review, the UM reviewer clicks 📷 Read this image with Verum · Vision. Findings stream into a right-side drawer in 3-8 seconds. The reviewer attests the relevant ones and clicks Promote. The deterministic gate re-derives against the augmented fact bank — the BRIDGE-extended rules now fire MET — and the verdict flips PEND → APPROVE.
The downstream artifacts inherit the evidence. The appeal letter generator and the SOAP note generator both cite the AI-attested image findings inline, with full attribution — reviewer name, vision-language model, timestamp, verbatim visual evidence quote. The decision-packet PDF gets a Page-1 banner and a Page-3 audit section. Append-only on the Decision: every promotion is reversible via the existing revert chain.
BRIDGE-extended rules carry a brass “BRIDGE” pill in the cited-rules list — so a reviewer can always tell which rules derive from core Medicare / CMS policy versus the additive AI-evidence ontology layer.
- 01READPEND
Verum · Vision reads the attached image
MedGemma 4B foundation, fine-tuned for our 4 imaging modalities. 3-8 seconds warm. Emits structured findings keyed to the policy vocabulary.
- 02ATTESTPEND
Reviewer attests + promotes
High-confidence findings pre-checked; reviewer toggles + clicks Promote. Audit captures the reviewer’s name, timestamp, model, verbatim visual evidence.
- 03RE-DERIVEAPPROVE
Gate re-evaluates · verdict flips
BRIDGE-extended rules fire MET against the augmented fact bank. Deterministic gate re-derives in pure code. Appeal letter + SOAP note auto-cite the AI evidence with attribution.
sample · bridge-lumbar-decompression-imaging-gap