Two-Layer Shadow-Auditor Pre-Submission Denial Predictor

Executive Summary

JIL Sovereign's shadow auditor predicts, before a claim is ever submitted, whether a government or payer auditor would deny it. The prediction is produced by two distinct layers. A deterministic rule layer evaluates the claim against the compliance canon: if any rule triggers, the deterministic layer's prediction is authoritative - "would deny" - full stop. A subordinate, self-calibrating advisory layer computes, for each triggered rule code, a denial probability derived from a Laplace-smoothed, sample-size-weighted blend of that specific engagement's own historically realized outcomes and a catalog-level severity prior, then combines the probabilities across all triggered codes on the claim using a noisy-OR function.

The result is a prediction that is both explainable - an auditor can trace the authoritative call to the specific rule that triggered it - and nuanced, because the advisory layer surfaces relative risk across multiple weaker signals that the deterministic layer alone would treat as a flat binary.

Problem Statement

Predictive denial-risk tools used in healthcare revenue cycle are typically pure black-box machine-learning models trained on aggregate historical data across many providers. Two limitations follow from that design. First, a pure ML score cannot explain why a specific claim is predicted to be denied in terms an auditor would actually cite - a probability number alone gives a provider nothing actionable to fix. Second, treating every provider's history as interchangeable ignores that a specific engagement's own track record with a specific payer or auditor is itself informative, and that history changes as more of that engagement's own claims are adjudicated over time.

Why Existing Approaches Are Insufficient

• Pure black-box ML denial-prediction tools: Output a single probability with no traceable, citable reason and no guaranteed authoritative floor - a false negative is indistinguishable from a false positive to the end user.
• Pure deterministic rule engines with no probability layer: Only flag a binary trigger/no-trigger outcome and provide no way to rank relative risk across multiple weaker, non-triggering signals present on a claim.
• Pooled-population-only models: Do not calibrate to a specific engagement's own realized outcome history, missing information that becomes increasingly informative as an engagement accumulates its own appeal record.

Technical Architecture

Per-Code Probability Calibration

For each triggered rule code, the advisory layer computes a Laplace-smoothed probability blending that specific engagement's own realized denial outcomes for that code with a catalog-level severity prior, weighted by the sample size of realized outcomes available - so a new engagement with little history leans on the catalog prior, and a mature engagement's estimate shifts toward its own experience as more outcomes accumulate.

Noisy-OR Aggregation

Per-code probabilities for all rule codes triggered on a given claim are combined via a noisy-OR function, producing a single aggregate advisory probability that increases when any strong individual signal is present without simply summing probabilities past 100%.

Noisy-OR Combination and Calibration

Noisy-OR reflects the intuition that multiple independent risk signals should combine such that any one signal strong enough on its own drives the aggregate estimate up, without the double-counting distortion of naive addition. Laplace smoothing prevents a rule code with few realized outcomes early in an engagement's history from collapsing to an unstable 0% or 100% probability. As an engagement accumulates its own realized outcomes for a given code, the sample-size weighting progressively shifts the blended estimate away from the catalog-wide prior and toward that engagement's own experience.

Prior Art Differentiation

Patent Claim