Exam pack · NAIC AI Risk Evaluation Supplement v6.0 (exposure draft, 16 Oct 2026); v7.0 pending adoption vote, Fall National Meeting

Bluebonnet Mutual Insurance Company (fictional)

NAIC code DEMO-10101 · period 2026-07-02 to 2026-10-31 · all states · built 2026-10-30 11:45

Pack identifier: pack_e4072ef8b88d
Pack hash (on ledger #1375): c56fad6b39de203033bd60e843014f878c2cefc5e1ac9c22ef9f6c754297a454
Engine: regbench-engine/0.4.0 · rules 2026.10+dbe5ddf6 · corpus 2026-10-30-5f07c521
Totals: 577 assessments · 547 case files · 4 high-risk models
Confidentiality: Produced for the requesting state's examination or inquiry; to be held confidential under that state's authority. The carrier may assert trade-secret protection over model documentation it attaches.

Download JSON Print

Exhibit A · use of AI Exhibit B · AIS Program Exhibit C · high-risk models Exhibit D · model data RegBench disclosure Ledger

Third-party AI system disclosure: RegBench (Exhibit C format)

The Supplement holds the carrier responsible for third-party AI systems it relies on. RegBench answers Exhibit C about itself so the carrier does not have to.

1 name version	RegBench assessment engine regbench-engine/0.4.0; rules 2026.10+dbe5ddf6; corpus 2026-10-30-5f07c521
2 use cases	Regulatory assessment of AI-influenced insurance decisions; evidence assembly for examinations. Decision support only; the carrier remains the decision-maker.
3 model type	Deterministic rules engine (rules as data). No machine-learning model produces findings. Optional narrative polish by a pinned language model is disabled in this configuration.
3a agentic	agentic: False
4 implementation date	None
5 development	vendor: third_party vendor name: RegBench (Kenbright Actuarial and Financial Services North America Inc)
6 purchased modifiable	purchased: True developed for company: False modifiable: doctrine-layer findings configurable as advisory; black-letter rules not modifiable by the carrier
7 model risks	False negatives where a jurisdiction, line or stage is outside rule coverage (see coverage map); false positives on doctrine-layer findings (advisory).
8 limitations	Assesses recorded decision facts, not model internals; coverage limited to the states, lines and stages in the rule pack; provisions marked 'verification pending' are drafts.
9 automate augment support	support
10 testing validation	Rule-level unit tests with positive and negative cases; citation integrity fails closed (100% existence check); point-in-time tests; eval set of showcase cases.
10a proxy evaluation	RegBench does not infer demographic characteristics. Portfolio tests consume carrier-produced proxy labels inside the carrier environment and release aggregates only.
11 last test date	2026-10-30
12 financial statement effect	None (no financial transactions generated).
13 legal compliance review	Rule pack reviewed against primary sources with verification status recorded per provision.
14 regulatory actions	None.