What this demo is, and isn't

• All patient names, member IDs, claim IDs, and denial records are synthetic. CARC and RARC codes are real WPC ANSI X12 835 values; CPT and HCPCS codes are real. The data behind every code is fictional; no PHI, no real denials, no real appeal letters.
• The triage classifier is the economic gatekeeper. Its outputs are pre-computed in the synthetic dataset. A production deployment would call a gradient-boosted classifier or a fine-tuned distilbert with features like CARC + RARC + payer + CPT + dollar amount + days since denial + prior appeal count + clinical doc availability. The decision shape (recover / borderline / write-off), the confidence score, the per-feature contribution breakdown, and the expected-economics arithmetic are all modeled on what a production classifier would emit — not narrative text dressed up to look like a model output.
• The appeal drafter is the headline agentic surface, and the drafted letters in the synthetic dataset are pre-computed to keep the demo deterministic and fast. A production deployment would call a fine-tuned LLM (Claude or GPT) with the relevant template + denial provenance + payer policy corpus + clinical context as input, and emit the same output shape (letter text + citations array + supporting docs list + word count). The template structure is the durable contract that the rest of the system builds against; the LLM output is the variable piece.
• Citation references in the drafted letters use synthetic policy numbers and section references. The format of the citation (e.g. Payer Medical Policy MP-{policy_number} Section {section}) is real; the specific MP- numbers and section references are synthetic. In a real deployment the drafter would resolve actual payer policy URLs and section anchors from a maintained corpus.
• The outcome predictor models a survival-analysis classifier. It outputs overturn probability with a confidence interval AND expected days-to-resolution with a range — a joint distribution. Production would fit this against a labeled corpus of historical appeals per payer / per CARC / per template. The CI widths in the synthetic dataset are illustrative; real CIs would be calibrated against historical model error.
• The 5 appeal templates are illustrative, not authoritative. Real RCM organizations maintain template libraries that are frequently larger (often 15-30 named templates) and are tuned to each payer's preferred appeal format. The template structure shown here — required sections, citation format spec, supporting-doc checklist — is the right shape; the specific 5 templates are common archetypes selected for demo clarity.
• Economics is grounded in plausible real-world ratios. The 10-20x cost difference between front-end correction and downstream appeal is real and durable. The write-off threshold shift from $200-500 to $25-75 is a function of how much labor the drafter removes from the appeal-and-submit workflow. The $1M-$3M annual additional recoverable assumes a mid-size health system; smaller organizations would see proportionally less. Real organizations vary widely — calibrate the absolute dollars to the specific deployment.
• The risk patterns and CARC mappings in this dataset reference taxonomy node IDs (e.g. tax-medical-necessity-denial) for narrative consistency with the denial-spine demo. The IDs are referenced as plain strings here — there's no runtime join across demos. In a real deployment the taxonomy is shared infrastructure; here the linkage is illustrative.
• The recover-rate metric (~65% in this dataset) is a function of how the synthetic denials were sampled. A real deployment's recover-rate would land lower if the underlying payer mix is heavy on contractual write-offs or patient-responsibility, higher if the payer mix is heavy on medical-necessity and auth denials. The decision logic is the durable piece; the rate is a function of the input distribution.