Insurance Book-Roll Panel Kit
Free for noncommercial use under PolyForm Noncommercial 1.0.0; commercial use needs a licence from Rico. View the source
The agency's letter goes in. Thirteen dossier-backed policyholder seats read it independently and say whether they would call back, stay, or start shopping. A weighted rubric scores the letter, the reactions are synthesized without erasing dissent, and three rewrites are tested the same way. The agency's producers and account managers keep every decision; the panel gives them the objections early.
Experimental
The problem
A renewal or book-roll letter goes to the whole book at once. Finding out how it landed by mailing it is the expensive way, and the feedback arrives as non-renewals. Testing it properly first (recruiting policyholders in the right lines, moderating, writing it up) is a research project no small agency commissions before a mailing.
How it works
The kit is a re-skin of a general synthetic customer panel: the runner, scoring engine, rubric machinery, seat contracts and validators are reused, and only the insurance layer (seat shapes, letter-input rules, workflows) is new. A deterministic TypeScript gate checks the input before any AI work starts. Each seat runs in its own fresh context against the same letter, then a synthesis step and a frozen scorecard combine the reactions. room.sh manages named rooms: build the panel once from the agency's book, then clean resets it for the next letter, archiving rather than deleting.
Features
- A script-enforced refusal
- Own-data verification
- Honest execution records
- Harness-agnostic tiers
- Room lifecycle in plain bash 3.2
- Worked examples
How it was built
Version 0.1.0 was built on 2026-08-12 for independent insurance agencies that send renewal and book-roll letters to the whole book and only learn the reaction from a wave of calls.
An AI agent built it as a re-skin of the existing customer panel kit. The runner, scoring, validators and tests were reused unchanged, and only the insurance layer was written new.
It shipped with thirteen policyholder seat shapes, two deliberate dissent seats and refusals for a missing letter, a missing book profile or an unexplained premium change. Every behavioural claim is labelled a modelled assumption until checked against a real mailing.
The same day AI agents ran it as three invented agencies. The premium-change refusal missed softened wording it existed to catch, so version 0.1.1 fixed it.
Rico kept the agency in charge of every coverage, pricing and compliance decision; the kit makes none of them.