PolicyChat Verdict Engine — Decision Recommendation Methodology
PolicyChat Verdict Engine — Decision Recommendation Methodology
Effective: 2026. Maintained by: PolicyChat Editorial.
The PolicyChat Verdict Engine produces 30-second specific recommendations for common insurance decisions: auto liability limits, term vs whole life, umbrella coverage need, comprehensive vs collision tradeoffs. Each recommendation is gated by an explicit conviction tier; the engine refuses to publish magnitude claims when its confidence is below the validated threshold.
1. Decision-type scope
The Verdict Engine currently produces recommendations for:
- Auto liability limits: $100K/$300K vs $300K/$300K vs $500K/$500K
(with optional umbrella). The decision rule anchors on
max(annual income × 1.5, asset value × 1.5). - Term vs whole life: anchored on dependent count, household income, debt load, and tax-advantaged-account utilization status.
- Umbrella need: above $500K asset value OR high-risk factors (pool, teen driver, multiple properties), recommend $1M umbrella requiring $250K/$500K underlying auto.
- Carrier-specific recommendations: directional only, never magnitude (see Section 4).
2. The conviction-tier framework
Every recommendation surfaces a confidence tier:
- Validated: n≥30 historical observations AND posterior CI lower bound exceeds the runner-up’s CI upper bound. Published with specific dollar magnitudes.
- Directional only: n≥15 AND posterior point estimate beats runner-up by ≥0.10 absolute. Published as a ranking (e.g. “X beats Y”) without dollar magnitudes.
- Kill-log: insufficient data or runtime model edge below the directional threshold. The engine refuses to publish a recommendation; Sage hands the user a conversational fallback.
This framework derives from chorus_stage2’s |p−0.5| > 0.2 publish gate discipline. See /methodology/conviction-tier/.
3. How recommendations are derived
For each decision type, the engine evaluates:
- The applicant profile (age, household income, asset value, dependents, driving record, etc.)
- A decision rule trained on a combination of:
- Industry-standard advisor guidelines (e.g. life insurance = 10× household income for working-age dependents)
- Empirical bounds from our Rate Authority ledger (typical premium deltas across coverage tiers)
- Asset-protection mathematics (umbrella threshold derivations)
- A competitive set of carriers known to write the recommended product class — sourced from the Rate Authority ledger.
4. Why carrier-specific recommendations are directional only
A specific “go with Carrier X at $Y/mo” recommendation requires a live quote pull — Sage performs this through the partner-feed adapter. The static Verdict Engine page surfaces RANKINGS (which carriers are typically competitive for the profile) but NEVER specific dollar magnitudes, because we cannot guarantee a bindable quote without running the partner-feed flow.
This is a deliberate design choice. Publishing a specific dollar figure that turns out to be wrong damages the citation authority that the rest of the methodology depends on. The Verdict Engine errs on the side of saying less when conviction is lower.
5. Refusal cases
When the engine returns a kill_log verdict, the user-facing page reads:
“PolicyChat doesn’t have validated conviction for a specific verdict on this question yet. Sage can walk through it conversationally and pull live quotes if you’d like.”
We do not pretend to have an answer the data does not support.
Maintained by PolicyChat Editorial. Methodology contact: editorial@policychat.com.