SSentry OS
How it works

Two stages. One verdict.
Under 200 ms.

Sentry runs every claim through a deterministic playbook engine first, then layers in Sonnet 4.5 reasoning for nuance. Both stages produce the same flag schema.

01
Ingest
Paste claim · CSV · PMS push.
02
Deterministic
NCCI · Mod · Freq · POS · Dx
03
AI reasoning
Sonnet 4.5 nuance pass
~5 ms<10 ms~1.8 s

Stage 1 — Deterministic playbook

Pure TypeScript. Zero I/O. Citation-perfect. These are the rules a senior biller would catch with a spreadsheet — except Sentry runs them in microseconds, every time, on every line.

  • NCCI Procedure-to-Procedure (PTP)
    97140↔97530, 97110↔97140, E/M↔procedure with modifier 25, plus the rest of NCCI Manual Ch. 11.
  • Modifier validity
    GP/GO required on PT/OT plan-of-care codes. 25 on same-DOS E/M+procedure. 59 only when 25/XE/XP/XS/XU don't fit.
  • Frequency / MUE
    Per-day medically unlikely edits per CMS MUE v2026. 4-unit ceilings on 97110/97140/97530.
  • Place of service
    POS 11/12/22/49 allowed for outpatient PT. POS 21 (inpatient) deflects.
  • Dx-procedure compatibility
    Functional-limitation diagnoses (M54/M62/M25/S/G ranges) for therapy CPTs per LCD L33631.
  • Payer-specific quirks
    Anthem CA aggressive 59 denial, Aetna retro-auth windows, UHC behavioral overlap.

Stage 2 — Sonnet 4.5 reasoning pass

Claude Sonnet 4.5 runs against a system prompt loaded with the full deterministic ruleset plus payer playbook intelligence. It returns ONLY structured JSON flags the deterministic layer didn't catch — never duplicates.

  • Medical-necessity narrative gaps
    Sentry asks: would a payer reviewer reject this dx-procedure pairing on subjective grounds even if rules pass?
  • Edge-case modifier combos
    XE/XP/XS/XU subset selection when 59 is overgeneral.
  • Coverage gotchas
    LCD coverage criteria not satisfied even when codes are valid.
  • Payer-specific subtleties
    Behavioral overlap on UHC, modifier-25 documentation calibration on Anthem.
  • JSON-only output
    Strictly schema-conformant. No prose. No hallucinated codes.
Architecture

Built on the family backbone.

Next.js 16 + Tailwind 4

Server-rendered React 19. Same stack as the rest of the family. Edge-deployed via Netlify.

Supabase Postgres

Shared family project. Auth, practices, scrub history. Row-Level Security on every table.

Anthropic Sonnet 4.5

$3/M input · $15/M output. Per-claim cost ~$0.005 at typical claim size. Monthly cap configurable.

Cross-product session

One signin works across ReimburseOS, Sentry, Denial OS, Reclaim OS, Leverage OS. Magic link or password.

BAA-eligible Anthropic workspace

PHI-handling claims are scrubbed under BAA. Patient identifiers are stripped before any model call.

Deterministic-first floor

AI failures never block scrubs. Stage 1 always runs. The floor of value is rule-based, citation-perfect.

Ready to scrub a real claim?

Start free. Five sample claims pre-loaded so you can see real flag output before you upload your own.

Get started The engine