4. Deterministic eligibility and rating

The model is useful before and after a rating decision: it gathers facts, explains a referral, and helps a customer choose. It must not decide whether an application is eligible or manufacture money. HomeInsuranceQuoteFlow makes that boundary a model-free LogicStep and a plain TypeScript rating service.

The goal

• Load a local rules file as an explicit, validated dependency.
• Build a quote from the four authoritative input states.
• Route an eligible quote differently from a referral or unsupported location.
• Make deterministic outputs easy to test without a provider.

The LogicStep boundary

RateQuoteStep does not call a model. It reads one state record from each input owner, calls RatingEngine.quote, saves the result, and routes:

public async runLogic(): Promise<LogicResponseType> {
  const application: QuoteApplication = {
    qualification: this.requireState<Qualification>(QualificationStep, "qualification"),
    property: this.requireState<PropertyProfile>(PropertyStep, "property"),
    risk: this.requireState<RiskProfile>(RiskStep, "risk"),
    coverage: this.requireState<CoveragePreferences>(CoverageStep, "coverage"),
  };
  const quoteResult = RatingEngine.quote(application, homeInsuranceCurrentDate());
  this.saveState({ quoteResult: durableJson(quoteResult) });
  if (quoteResult.decision === "eligible") {
    return go(PresentQuoteStep).withState({ needsPresentation: true });
  }
  return go(ReferralStep).withState({
    decision: quoteResult.decision,
    reasonCodes: quoteResult.reasonCodes,
  });
}

requireState(...) turns a missing prerequisite into a programmer-visible error; it does not ask the model to guess. A quote is calculated only after the review tool has confirmed a complete application.

Versioned rules, explicit outcomes

quote-config.json is a deliberately fictional product configuration. Its Zod loader validates supported states, limits, factors, tiers, endorsement fees, referral thresholds, and a rulesVersion at module load. RatingEngine.quote returns exactly one of these shapes:

The order matters: an unsupported state is not a partially priced referral. It has no generated option table at all.

Money and identity belong in the engine

The calculation uses typed factors and rounds in one place. The quote identifier is a hash of the input application plus the frozen current date:

private static money(value: number): number {
  return Math.round((value + Number.EPSILON) * 100) / 100;
}

private static quoteId(application: QuoteApplication, currentDate: Date): string {
  const digest = createHash("sha256")
    .update(JSON.stringify(application))
    .digest("hex")
    .slice(0, 8)
    .toUpperCase();
  return `EHI-${currentDate.toISOString().slice(0, 10).replaceAll("-", "")}-${digest}`;
}

Changing the deductible changes the serialized input, so it changes both the premium and quote ID. The model can describe that fact, but it cannot choose a different price or continue presenting an obsolete selection.

Test the boundary without an LLM

The end-to-end test also contains deterministic service tests. They prove that the same application yields the expected option IDs and premiums, that an unsupported state produces no options, that an old roof produces a referral, and that a deductible change produces a lower premium and new quote ID.

The live scenario uses credentials from .env and is skipped when they are unavailable.

This is the fast check for code-owned policy. The live replay in lesson 7 adds the separate question of whether the model follows the conversation contract.

Next

5. Memory namespaces and expiry explains why the three long intake stages share history while rating has none.