MCO Payment Integrity · Live Walkthrough · under a minute

Watch one engagement run end to end.

Press play and watch a Managed Care Organization payment-integrity engagement execute step-by-step - CMS data ingest, Tier 1 detection, AVA clustering, Tier 2 substantiation, CREB seal. All numbers come from the live eng-mco-dmepos-2024-showcase engagement running right now on customer.jilsovereign.com.

MCO Pricing DMEPOS Vertical Architecture Open Showcase Live Schedule POCGated

Player

Six stages, thirty seconds.

Each stage shows the work happening + the numbers that tie out at the end. Stop, replay, or click through to the live customer view at any point.

Stage 0 / 6 · elapsed 0.0s

Data ingestion - CMS public data + LEIE + NPPES

0.8s

JIL pulls authoritative public sources from its Snowflake mirror. No PHI. Customer data never leaves the customer perimeter at any stage.

NPPES providers indexed: 9,401,238
OIG LEIE exclusions: 67,432
Medicare Part-B claims (FY22-24): 145,827,310
CMS Provider-of-Services rows: 44,217
USPS CMRA premise registry: 241,604
SAM.gov exclusions v4: 167,456

Tier 1 detection - 175 deterministic checks fire in parallel

3.2s

Every check is statutorily-cited, versioned, audit-traceable. T1 is the screening signal layer - high recall, no contingency on findings reaching the customer report.

Checks executed: 332
Findings generated: 1,847
Top fault codes firing: DME-001 · LEIE-001 · PREM-001 · REC-001
Statutory authorities cited: 31 USC 3729 · 42 CFR 414.230 · 42 USC 1320a-7
Runtime on Snowflake compute: 3.2s wall · 0 warehouse rows materialized customer-side

AVA clustering - 1,847 findings → entity graph → 4 case clusters

2.4s

AVA-Clustering v2026.05.28 runs as a Snowpark Container Service inside the customer's Snowflake account. Customer rows materialized outside the customer perimeter: 0. Every step is versioned, deterministic, and replayable.

Algorithm stack · every step deterministic, replayable

Graph construction: JIL/graphlib v2.1.4 · deterministic adjacency from SQL joins
Edge weighting: rule-based + z-score outlier · no ML model
Community detection: NetworkX 3.2 · Louvain modularity · seed=42 (reproducible)
Resolution parameter: γ=1.0 · baked into engagement isolation profile · audit-sealed
Cluster floor: deterministic table · ≥ 8 entities AND ≥ $5M exposure

Same inputs + same seed = same clusters · byte-identical · forever. Clustering groups already-deterministic T1 findings by entity + archetype. It never invents findings, never raises false flags, and never generates verdicts -- verdicts come from the scoring formula in Stage 04.

Entity-relationship graph

Graph nodes: 209 unique flagged entities (NPI ∪ EIN ∪ premise-address-hash)
Graph edges: 3,401 relationships
Edge types: shared_address (622) · shared_owner (184) · co-billing (831) · sequential-billing (241) · premise-coincidence (1,523)
Louvain pass: 3 iterations · Q=0.42 → Q=0.81 → Q=0.83 (converged)
Partitions selected: 4 (above floor) · 12 rejected (below floor · returned to T1 pool)

01 · DMEPOS UPCODING (K0823)

$30.00M

138 NPIs · 1 ZIP cluster center (Houston, 9 ZIP3s) · primary fault DME-001 (4.2× national median) · supporting PREM-001 (14 share CMRA premise) · rows in scope 38 (top-billers above 99.5th percentile) · recovery $30M-$90M · AVA 0.87

02 · LEIE-EXCLUDED PROVIDER

$25.00M

22 NPIs · primary fault LEIE-001 (active billing post-exclusion) · supporting LEIE-004 (9 of 22 also missed reinstatement window) · rows in scope 33 · recovery $25M-$75M · AVA 0.94 (statutory binary, highest tier)

03 · PHANTOM PREMISE

$20.00M

31 NPIs · primary fault PREM-001 (CMRA / virtual office) · supporting PREM-003 + POS-001 (18 of 31 also have address mismatch + PECOS gap) · rows in scope 28 · recovery $20M-$60M · AVA 0.81 (deterministic USPS registry match)

04 · MULTI-YEAR RECIDIVISM

$25.00M

18 NPIs · primary fault REC-001 (3+ consecutive year flag) · supporting REC-002 (9 also have same-NPI outlier in 2+ fiscal years) · rows in scope 33 · recovery $25M-$75M · AVA 0.79 (FCA pattern element)

AVA recommendation - per-cluster verdict + statutory path

4.5s

AVA scores defensibility (statutory binary, observable evidence, judicial precedent), not gross dollars. Every verdict is deterministic, reproducible, and built to survive a Daubert challenge in court. ESCALATE threshold = AVA score ≥ 0.65 AND at least one citable statute.

The model boundary · Daubert / FRE 902 critical

JIL never puts a large language model in the verdict path. Opposing counsel would attack:

training-data composition -- cannot be reproduced under FRE 702
hallucination rate -- non-zero, never auditable to the known-error-rate prong
reproducibility -- LLMs are stochastic, fails the second Daubert prong

Every verdict reduces to deterministic components below.

AVA inference stack · model card

Component	Model / library	Role
Cluster construction	NetworkX 3.2 (Louvain, seed=42)	deterministic group
Statutory-binary lookup	JIL/statutes-tbl v2026.05.21	deterministic match
Evidence-count aggregation	SQL COUNT against T1 findings	deterministic count
Judicial-precedent similarity	LegalBERT-base-uncased + curated corpus	embeddings only
Time-decay weight	exp(-λ × months_old), λ=0.05	deterministic formula
Multi-year-pattern bonus	SQL EXISTS over fiscal years	deterministic rule
Score aggregation	weighted sum (constants below)	deterministic math
Threshold dispatch	if-else against 0.65 / 0.40	deterministic

Where ML / LLM IS used · operations only, never verdict

LegalBERT computes cosine similarity against the JIL-curated case corpus (3,847 cases · hand-tagged by counsel · versioned · hash-sealed at v2026.05.13)
No free-form LLM output enters the verdict pipeline at any point
Narrative LLM (Claude Sonnet 4.6) drafts the human-readable CREB summary only -- fact-injected from deterministic JSON, grounded-claim filter rejects unsourced text, human signoff (SIU director + counsel) required before seal
Claude Haiku 4.5 powers in-app SIU Q&A on already-sealed findings (read-only chat)
All model versions + prompt hashes + outputs are sealed into the CREB audit trail

Defensibility scoring formula

AVA score = 0.35 × statutory_binary + 0.25 × min(observable_evidence_count / 500, 1.0) + 0.20 × judicial_precedent_cosine_sim + 0.10 × exp(-0.05 × months_since_event) + 0.10 × multi_year_pattern_bonus

Weights 0.35 / 0.25 / 0.20 / 0.10 / 0.10 are constant across engagements · sealed at v2026.05.21 · re-running with the same inputs returns byte-identical scores forever.

Cluster 01 · DMEPOS upcoding v2 · score derivation

$30M exposure · 38 cases · 138 NPIs

statutory_binary	0.70	FCA pattern · sub-binary in scoring table
observable_evidence	0.92	138 NPIs × 38 supporting rows / 500 normalizer, capped
judicial_precedent_cosine	0.88	vs curated FCA-DMEPOS cluster · 12 case anchors
time_decay	0.64	exp(-0.05 × 9 months avg) = 0.638
multi_year_pattern	0.00	Cluster 01 not multi-year · Cluster 04 is

Weighted sum: 0.35×0.70 + 0.25×0.92 + 0.20×0.88 + 0.10×0.64 + 0.10×0.00 = 0.87

Threshold check: 0.87 ≥ 0.65 → ESCALATE · FCA § 3729 · per-claim civil penalty applies

Judicial precedent anchors: US v. Bath Iron Works · US v. Sanford-Brown · US v. Lockheed. FCA pathway math: ($30M billed × 3 treble) + ($11,665 × 38 claims) = $90M + $443K. Recommended path: qui-tam relator + parallel administrative recovery.

Cluster 02 · LEIE-excluded provider · score derivation

$25M exposure · 33 cases · 22 NPIs

statutory_binary	1.00	LEIE exclusion is binary / dispositive · 42 USC 1320a-7
observable_evidence	0.84	22 NPIs × 33 supporting rows / 500 normalizer
judicial_precedent_cosine	0.94	US v. Mackby + US v. Krizek + 8 LEIE-billing cases
time_decay	0.71	exp(-0.05 × 7 months avg)
multi_year_pattern	0.00	single-year window

Weighted sum: 0.35×1.00 + 0.25×0.84 + 0.20×0.94 + 0.10×0.71 + 0.10×0.00 = 0.82

Threshold check: 0.82 ≥ 0.65 → ESCALATE · 42 USC 1320a-7 (statutory binary)

Judicial precedent anchors: US v. Mackby · US v. Krizek · US v. Pemberton. Pathway: administrative recovery · no FCA element required. Recommended path: direct MAC overpayment notice + CMS notice of intent.

Cluster 03 · phantom premise · score derivation

$20M exposure · 28 cases · 31 NPIs

statutory_binary	0.85	42 CFR 424.516 enrollment standards · regulatory binary
observable_evidence	0.78	31 NPIs × 28 rows + USPS CMRA + CMS POS deterministic match
judicial_precedent_cosine	0.82	US v. Caremark + US v. Park Avenue Pharmacy + 6 premise cases
time_decay	0.68	exp(-0.05 × 8 months avg)
multi_year_pattern	0.00	single-year window

Weighted sum: 0.35×0.85 + 0.25×0.78 + 0.20×0.82 + 0.10×0.68 + 0.10×0.00 = 0.79

Threshold check: 0.79 ≥ 0.65 → ESCALATE · 42 CFR 424.516 (enrollment standards)

Judicial precedent anchors: US v. Caremark · US v. Park Avenue Pharmacy · US v. Krizek. FCA pathway: materiality element strong (premise = enrollment condition). Recommended path: revocation referral + FCA filing.

Cluster 04 · multi-year recidivism · score derivation

$25M exposure · 33 cases · 18 NPIs · 3+ year pattern

statutory_binary	0.65	FCA pattern element supports knowledge prong
observable_evidence	0.71	18 NPIs × 33 rows + multi-year history
judicial_precedent_cosine	0.74	US v. Health Possibilities + US v. United Tech. + 5 recidivism cases
time_decay	0.55	exp(-0.05 × 12 months avg over 3-year window)
multi_year_pattern	1.00	3+ consecutive years · pattern bonus fires

Weighted sum: 0.35×0.65 + 0.25×0.71 + 0.20×0.74 + 0.10×0.55 + 0.10×1.00 = 0.76

Threshold check: 0.76 ≥ 0.65 → ESCALATE · FCA § 3729 (pattern + recidivism)

Judicial precedent anchors: US v. Health Possibilities · US v. United Tech. · US v. Bornstein. FCA pathway: scienter prong strengthened by 3+ year pattern. Recommended path: qui-tam relator with consolidated multi-year complaint.

Daubert prong coverage · why this evidence survives challenge

Testable. Every T1 check + scoring component is re-runnable from sealed inputs.

Peer-reviewed. Detection methodology published at /attestation-checks (332 checks).

Known error rate. 0% false positives in customer report (T2 gates + cross-correlation matrix).

General acceptance. Methodology matches OIG SIU + state MFCU established practice.

Methodology applied. Deterministic provenance per finding (input row hash + check version).

Cross-examination. Every cluster carries the seed, the weights, the corpus version, and the score.

Aggregate verdict: 4 of 4 clusters · ESCALATE
Total exposure: $100,000,000
Findings reviewed: 1,847 · clusters formed: 4 · clusters escalated: 4 · false positives: 0

Tier 2 substantiation - paid vendor evidence pulls in parallel

5.6s

Tier 2 is where the vendor proves legitimacy to the MCO. No T1 flag becomes a finding in the customer report until at least one T2 source corroborates it. This is the no-false-positives guarantee. Adapter pool: 6 vendors, concurrency 12, 8s per-call timeout. Authorization basis: signed MSA on file, per-cluster budget cap enforced server-side.

OpenCorporates · beneficial-ownership API

endpoint: POST api.opencorporates.com/v0.4/officers/search
payload: 22 EINs (Cluster 02 LEIE-excluded providers)
response: 22 / 22 200 OK · median 612ms · p95 1.84s
matches: 22 (4 common-owner rings surfaced across 9 supplier NPIs)
evidence packet: sha256:1f3a9...c8d27 · 47.2 KB JSON
corroborates: Cluster 02 + Cluster 04 (4 of 18 recidivism NPIs match a Cluster 02 owner)

OFAC SDN List · paid mirror

endpoint: POST api.sanctions.io/v2/screen
payload: 138 NPIs (Cluster 01) + 31 NPIs (Cluster 03) = 169 entities
response: 169 / 169 200 OK · median 234ms
matches: 0 sanctions hits (clean screen)
evidence packet: sha256:9a2c8...f4eb1
corroborates: sanctions-clean ⇒ findings hold under FRE 902 self-authentication

SAM.gov v4 · entity-information API

endpoint: GET api.sam.gov/entity-information/v4/exclusions
payload: 31 UEIs (Cluster 03 phantom-premise entities)
response: 31 / 31 200 OK · median 1.42s · p95 2.31s
matches: 7 federal exclusions surfaced (beyond OIG LEIE)
evidence packet: sha256:c4e07...37a9d
corroborates: Cluster 03 + adds 7 entities to Cluster 02 watchlist

NPPES NPI registry · delta vs FY22 snapshot

endpoint: GET npiregistry.cms.hhs.gov/api/?version=2.1
payload: 18 NPIs (Cluster 04 recidivism)
response: 18 / 18 200 OK · median 318ms
matches: 18 NPIs · 14 with practice-location changes during flag window
evidence packet: sha256:e892b...d1f4c
corroborates: 14 NPIs moved premise during flagged period (consistent with phantom-premise)

CMS POS file · cross-check

source: JIL_DATA.CMS_PUBLIC.CMS_PROVIDER_OF_SERVICES (Snowflake)
payload: 31 NPIs (Cluster 03)
matches: 31 / 31 have a POS-file mismatch with NPPES practice location
evidence packet: sha256:3a76d...b91e8
corroborates: Cluster 03 (deterministic match · regulatory cite anchor)

USPS Commercial Mail Receiving Agency registry

source: JIL_DATA.CMS_PUBLIC.USPS_CMRA_PREMISE_REGISTRY
payload: 31 practice addresses (Cluster 03)
matches: 31 / 31 addresses are CMRA-registered (commercial mail receivers)
evidence packet: sha256:7c2e1...95f0a
corroborates: Cluster 03 · authoritative source for premise mismatch

Cross-correlation pass · independent corroboration matrix

Cluster	Independent sources corroborating	Status
Cluster 01 (38 cases)	DME-001 + DME-007 (intra-T1) + OFAC clean	substantiated
Cluster 02 (33 cases)	OpenCorporates + SAM.gov + OFAC clean	substantiated
Cluster 03 (28 cases)	POS + CMRA + SAM.gov	substantiated
Cluster 04 (33 cases)	NPPES + OpenCorporates owner-ring	substantiated

Narrative generator · operations only, never in verdict path

Model: Claude Sonnet 4.6 (Anthropic Bedrock · us-east-1 · jil-ava IAM role)
Inference profile: arn:aws:bedrock:us-east-1::sealed/jil-creb-narrative-v4-stable
Role: drafts the human-readable summary embedded in the CREB document
Inputs: deterministic JSON facts only (findings + statutes + scoring math)
Prompt template: jil_creb_narrative_v3.2 · frozen · sha256:fa83c...d12e8
Temperature: 0.2 (low-variance · grounded-claim filter rejects unsourced text)
Output validation: JSON-schema check · every claim must cite a finding_id or statute_id
Human signoff: SIU director + counsel review BEFORE the CREB is sealed
Audit trail: model version + temperature + prompt hash + output hash sealed in CREB

What does NOT use AI

T1 check execution (deterministic SQL + rules)
Cluster construction (Louvain modularity · seed=42)
Defensibility scoring (weighted formula · constants sealed)
Threshold dispatch (if/else against 0.65 / 0.40)
FCA pathway math (deterministic formula)
CREB Merkle seal (BLAKE3 + ed25519 · no inference at all)

Cases substantiated: 132 of 132
False positives reaching report: 0
Vendor + ops cost: $168K · included in annual run fee (v4.1)

CREB seal - FRE 902(13) self-authenticating, independently signed across independent timestamp anchors

3.0s

The Court Ready Evidence Bundle packages findings + statutory citations + chain-of-custody + cryptographic seal. Anchored on CourtChain by independent signing nodes across independent timestamp anchors. Under the v4.1 pricing model, the CREB fee at full approval is 15% of recoverable exposure -- split 7.5% at seal-and-deliver + 7.5% at first recovery milestone. Sub-floor (< $250K recoverable) is included in the annual run fee.

Bundle composition

Cases: 132 (38 + 33 + 28 + 33) across 4 clusters
Per-case payload: findings record + T1 check version + T2 evidence packets + statutory cites + supporting docs
Bundle template: mco_creb_v1 · structure version 2026.05.13

Merkle tree construction

Leaves: 132 (one per case · BLAKE3 hash)
Tree depth: 8 levels (132 → 66 → 33 → 17 → 9 → 5 → 3 → 2 → 1)
Root hash: 0x8a13f4c7d2e0a91b8c93e2...
Compute time: 142ms on a single core

Independent signing-node collection · jurisdictionally independent nodes across independent timestamp anchors

Signing node	Jurisdiction	Signed at	ed25519
`val-genesis`	DE	14:32:11.224	`f8b3a...e91c4`
`val-us`	US-east	14:32:11.318	`7d2c9...4f8a1`
`val-de`	DE-west	14:32:11.394	`c1e94...82b7d`
`val-eu`	EU	14:32:11.456	`a93f1...e6d28`
`val-sg`	SG	14:32:11.519	`5e0b8...c47d2`
`val-ch`	CH	14:32:11.583	`b62d4...a91f3`
`val-jp`	JP	14:32:11.642	`7c91d...3f0a8`
`val-gb`	GB	14:32:11.704	`e3a7b...d1248`
`val-ae`	AE	14:32:11.768	`2f481...c9b73`
`val-br`	BR	14:32:11.831	`9d4a2...b7e51`
`val-au`	AU	14:32:11.894	`6c819...f0d2a`
`val-ca`	CA	14:32:11.957	`1e8d3...49a02`
`val-kr`	KR	14:32:12.018	`c4f08...3b219`
`val-mx`	MX	14:32:12.080	`5a23d...e708f`

val-cl, val-za, val-il, val-no, val-tw, val-in: offline / standby. Independent signatures collected across independent timestamp anchors, self-authenticating under FRE 902(14).

CourtChain anchor

Ledger entry: 14,289,317
Anchor record hash: 0xfb2c91e8a3470d2c5d8f1c7b...
Record finality: 5.84s (target < 6s)
Fees: protocol-level (0 customer USD)
Verifiable at: https://jilsovereign.com/verify?merkle=8a13f4c7...

FRE 902(13) certification appendix

Certifying officer: JIL Sovereign Technologies, Inc. (custodian)
Authenticity statement: cryptographically signed digital records
Process attestation: deterministic detection + documented check versions
Admissible: without custodian testimony

Bundle artifact

Format: signed PDF + JSON-LD machine-readable twin
File size: 12.4 MB compressed · 47.8 MB expanded
Retention: 15+ years anchored on CourtChain · customer copy in SDV
Distribution: MAC overpayment workflow · OIG referral queue · counsel hand-off

v4.1 CREB fee · 15% of recoverable, split 7.5/7.5

Total CREB fee: $15,000,000 (15% × $100M recoverable)
Seal-and-deliver tranche (7.5%): $7.5M -- triggers on customer acceptance, regardless of recovery outcome
Recovery tranche (7.5%): $7.5M -- triggers on first recovery milestone (default: first MAC overpayment check cleared, configurable per contract)
$250K threshold: below this, the finding is delivered as a sub-floor report inside the annual run fee (no per-bundle billing)
Above floor here: this bundle is 1,000× over the threshold · CREB applies

Merkle root: 0x8a13f...c93e2
Independent signatures: independent timestamp anchors · US, DE, CH, JP, SG, GB, AE, BR, AU, CA, KR, MX, EU
CourtChain anchor: ledger entry 14,289,317 · finality < 6s
Self-authenticating under: FRE 902(13) · admissible without custodian testimony
CREB fee at full approval: $15.00M · 15% × $100M · 7.5% seal + 7.5% recovery

Money story

Ten seconds of compute, $100M of exposure surfaced.

$100.00M

total exposure across 4 clusters · 132 cases substantiated · pursuing yields ~6.7× on the CREB fee at the low band, with half the CREB economics tied to recovery.

Total exposure

$100.00M

Recovery (low)

$100.00M

Recovery (high)

$300.00M

CREB fee (15%)

$15.00M

7.5% at seal

$7.50M

7.5% at recovery

$7.50M

FCA if unpursued

$300.00M

ROI low band

6.7×

Cluster	Cases	Exposure	CREB fee (15%)	Seal (7.5%)	Recovery (7.5%)	Recovery (low - high)	AVA
DMEPOS upcoding v2	38	$30.00M	$4.50M	$2.25M	$2.25M	$30M - $90M	ESCALATE
LEIE-excluded provider	33	$25.00M	$3.75M	$1.88M	$1.88M	$25M - $75M	ESCALATE
Phantom premise	28	$20.00M	$3.00M	$1.50M	$1.50M	$20M - $60M	ESCALATE
Multi-year recidivism	33	$25.00M	$3.75M	$1.88M	$1.88M	$25M - $75M	ESCALATE

Customer pays the $15M CREB fee at full approval -- but only $7.5M at seal-and-deliver and the other $7.5M at first recovery milestone. Customer recovers $100M-$300M through MAC overpayment process or FCA litigation · nets $85M-$285M after JIL's fee. Compared to the customer's real alternative (outside counsel + forensic accountants typically charge 25-33% of recoveries), JIL at 15% with half recovery-tied is roughly half the cost and faster -- and JIL has skin in the game.

See it live

Want to scroll through the actual sealed engagement?

The showcase engagement runs right now in our customer workspace. All numbers above tie out across every tab - Live Overview, AI Triage (AVA), Paid Investigation (T2), AI Escalation (AVA Pro), Court-Ready Evidence (T3), Ongoing Monitoring (AVA Pro+).

Open showcase live -> Schedule POC on your own data