VIN-repeat, geo-cluster, and fatal-rate outliers from real NHTSA crash data.
pc-engine ingests the NHTSA FARS public crash register, indexes by VIN, lat/lon, and make/model/year cohort, and runs three deterministic checks per engagement. Same kernel that ships every other JIL vertical: customer-profile gated on lob = 'pc_insurance_carrier', sealed CREB on every finding, FRE 902(14) self-authentication. One kernel, 8 industry verticals, 175 production checks, 273 production services.
What this POC shows.
If you're a P&C insurance SIU, a state insurance regulator, or a reinsurer, this is the short answer for what's actually being detected here.
What's the dataset?
NHTSA FARS (Fatality Analysis Reporting System) public crash data. 60K+ records. Public, free, deterministic. Same data your in-house actuarial team has access to.
What did JIL find?
VIN-repeat patterns (same VIN appearing in multiple high-cost claims), geographic clusters of fatal-rate outliers, NAICS-volume anomalies in commercial-fleet incidents. 450 findings ranked by anomaly score with the contributing data + threshold.
Is this useful for SIU?
Yes for first-pass triage. Frees your investigators from manually ranking the noise pool. Each finding has the deterministic rule that fired + the contributing data row, so your SIU can investigate with the methodology in hand.
What this is NOT
Not a fraud determination. Not a coverage decision. 'Flagged' means 'the data exhibits a pattern worth a human's time.' The verdict on whether it's fraud, error, or noise stays with your team.
How do I run this on my book?
Same engine, your loss-runs. We don't need PHI/PII; VIN + claim metadata + payment is enough. Turnaround: 7-14 days for first pass on a 24-month claim history.
What pc-engine fires on.
Same VIN in 2+ distinct crashes.
Same VIN appearing in 2+ distinct crash records. Cross-state span and presence of fatalities escalate severity. Leading indicator of staged-accident rings, VIN cloning, salvage-title fraud. Regulatory basis: NAIC Anti-Fraud Plan model, NHTSA defects investigation, TREAD Act 49 USC 30166.
Crashes clustered within 500m radius.
Crashes clustered within a configurable radius (default 500m) above a configurable threshold (default 5+ crashes). Material indicator of staged-accident rings operating from a single location. Regulatory basis: NAIC Anti-Fraud Plan model, NAIC P&C Annual Statement.
Make/model cohort 2.0x peer fatal rate.
Make/model/year cohort with fatal-crash rate at least 2.0x the peer mean for the same body-style and model-year cohort. Precursor signal NHTSA uses to open defects investigations. Regulatory basis: TREAD Act 49 USC 30166, SAE J2980.
What the carrier takes to NAIC or NHTSA.
finding_id : cf3b9210-...-pc-vin-repeat-001 check_id : pc_vin_repeat subject_type : vin subject_id : 1HGBH41JXMN000001 severity : critical occurrences : 5 distinct crash records state_count : 3 states fatal_count : 2 fatalities regulatory_basis: NAIC Anti-Fraud Plan, NHTSA defects investigation, TREAD Act 49 USC 30166 code_version : pc-engine@2026.05.01-pc-1
Deterministic, replayable, court-defensible.
Same kernel as the other 7 verticals. The VIN-repeat check is a SQL aggregate; geo-cluster bins crashes into a regular grid sized to the configured radius; fatal-rate outlier joins each cohort against the peer-mean for body-style + model-year. No stochastic LLM in the verdict path. Ava (the in-house agentic AI) groups, narrates, and routes findings; it never produces the underlying flag.
One kernel. Eight industries. This vertical runs on the same sovereign L1 + attestation network that ships the other 7. Kernel age: 18+ months. Adding a vertical: ~1 week. Competitor moat: build the kernel first.