Insurer and contributor pilot: Register an insurer, calibrate a contributor profile, or explore digital reviewers.
Three seats at the bench
The carrier
Sends each AI-influenced decision as recorded facts (never raw PII). Gets back a cited assessment, a review queue for FLAG and HOLD items, and an exam pack on demand.
The examiner
Receives packs structured as the Supplement's Exhibits A to D, verifies the hash chain independently, samples case files by model, and reads a cross-carrier benchmark that publishes only safe cells.
The carrier's agents and vendors
Claims, underwriting and pricing agents call RegBench through the Model Context Protocol before they act: check the action, record the decision, hand the case file to compliance.
Showcase cases
Fixed identifiers used in the walk-through. Each exercises a different regime: Texas prompt-payment and human-review expectations, Texas utilization review, Colorado's amended ECDIS regulation, New York's Circular Letter No. 7, the federal FCRA notice, and the NAIC bulletin where a state adopted it.
| Case | Carrier | State | Stage | Decision | Human review | Disposition | Most serious finding |
|---|---|---|---|---|---|---|---|
| TX-CLM-0417 | fnol-fraud v3.1 | TX | claims | deny | none | HOLD | critical Algorithmic denial, reduction or fraud hold with no human review |
| TX-UR-0052 | ur-auto v2.0 | TX | utilization review | adverse determination | nurse | HOLD | critical Adverse utilization-review determination without a physician's determination |
| CO-LIFE-0088 | rx-score v3 | CO | underwriting | decline | none | HOLD | critical Adverse action based on a consumer report with no FCRA notice recorded |
| NY-RATE-0231 | rx-score v3 | NY | rating | rate | underwriter | HOLD | high Adverse underwriting or pricing decision without specific reasons or disclosure of external data and AI use |
| CO-AUTO-0144 | price-opt v7.2 | CO | renewal | rate | none | FLAG | medium ECDIS used outside an attested governance and risk management framework (amended Regulation 10-1-1) |
| TX-UW-0311 | uw-triage v2.4 | TX | underwriting | approve | underwriter | PASS | none |
| CT-LIFE-0075 | rx-score v3 | CT | underwriting | decline | underwriter | FLAG | medium Adverse high-risk decision from a system with no adverse-outcome testing on record |
| TX-RNW-0199 | uw-triage v2.4 | TX | renewal | non renew | none | FLAG | high Consequential AI-supported decision taken without a person reviewing and agreeing |
Carriers in the demonstration book
| Carrier | Assessed | PASS | FLAG | HOLD | Reviews open | Median hrs |
|---|---|---|---|---|---|---|
| Bluebonnet Mutual Insurance Company (fictional) property & casualty mutual · TX, CO · auto, home | 593 | 477 | 95 | 21 | 12 | 24.8 |
| Harborline Health Plan, Inc. (fictional) health benefit plan issuer · TX, NY · health | 261 | 233 | 9 | 19 | 5 | 13.4 |
| Summit Range Life Assurance Company (fictional) life insurer · CO, TX, NY, CT · life | 283 | 139 | 123 | 21 | 27 | 30.2 |
What the engine knows, honestly
- Rules in force
- 30 rules in 7 files · version 2026.10+dbe5ddf6
- Corpus
- 39 provisions · 38 verified against primary sources · snapshot 2026-10-30-5f07c521
- Coverage gaps
- CO · fraud NY · claims NY · utilization review
No black-letter rule covers these cells yet (only advisory doctrine); decisions there pass by absence, and the pack says so. - Findings
- Deterministic. No language model produces a finding. Judgment-layer rules are advisory and can never suppress a black-letter finding.