The Engine — Sentry OS technical architecture

200ms

p50 latency

Deterministic stage. AI pass adds ~1.8s.

$0.005

per scrubbed claim

Sonnet 4.5 list price · ~1.5K input tok / 600 output tok.

99.7%

flag-citation accuracy

Citations are static — drawn from playbook, not generated.

6 stages

deterministic edits

NCCI · Mod · Freq · POS · Dx-Proc · Payer quirks.

CARC map

predicted denial code

Each flag maps to the CARC/RARC the payer would post (CO-97, CO-16, CO-50…).

JSON-only

AI output schema

Strict zod-validated. Failure → drop, not contaminate.

BAA-ready

Anthropic workspace

Patient ID stripped client-side before model call.

Why Sonnet 4.5 specifically

· 200K context — full payer playbook + NCCI ruleset fits in the system prompt.
· Tool-calling reliability — strict JSON output without retries.
· Cost — Opus 4.7 is >5x more expensive without measurable accuracy lift on this task.
· Latency — p50 ~1.8s vs Opus 4.7 ~5.4s. Throughput matters at PMS-batch volumes.

The system prompt

Compressed-but-complete encoding of the deterministic ruleset, the payer playbook library, and a strict output schema. Always asks the model to NOT duplicate flags already raised by stage 1, and to return an empty array on clean claims.

Schema:
{
  "flags": [
    {
      "severity": "BLOCKING" | "WARNING" | "INFO",
      "category": "NCCI-PTP" | "MOD-VAL" | "DX-PROC"
                | "POS" | "FREQ" | "PAYER-Q" | "MISSING-INFO",
      "code": string,
      "message": string,
      "citation": string,    // real reference, never fabricated
      "suggestedFix": string,
      "affectedLines": number[]
    }
  ]
}

Supabase shared backbone

practices, practice_members, profiles tables shared with sister products. sentry_scrub_results is product-scoped. RLS enforced on every read.

PHI handling

Patient ID and names stripped client-side BEFORE the model call. The scrubber sees CPT/dx/POS/payer/dates only — no PHI necessary for adjudication.

Failure mode

AI call fails → return deterministic flags only. AI call malformed → drop AI output, return deterministic. Never block on a soft signal.

Run a real claim through it.

Start free and scrub your first claim in minutes. Five sample claims pre-loaded so you can see flag output before you upload your own.

Try Sentry Pricing