Motor Claim Triage and QC Evaluator: Same Claim, Same Verdict, Every Time

Most fraud scores are sold as a black box, never checked against the insurer's own claims, and a probabilistic scorer will happily hand the same claim two different answers on two runs. This one is built the other way round.

Every claim is sorted into one of four categories, Fast track, Approve, Investigate, or Repudiate, by a fixed point score over the claim facts. There is no model in the classification path, so the same facts always score the same points and land in the same category, whether the claim is reviewed alone or inside the whole book. The two-answer problem is gone by construction.

The points come from what the data actually shows, not folklore. Across 15,420 real claims, the signals that separate fraud are policy-holder fault, an All Perils policy, a recent address change, and an accident at policy start. The intuitive ones fail here: claimants with no prior claims have the highest fraud, not the lowest, and 99% of claims are reported late, so "late report" tells you nothing. Those were dropped.

Then the scorecard. The fraud flags (Investigate and Repudiate) are measured against the real FraudFound_P label over the full population, so the ranges are tight, not a small-sample guess: catch rate, flag accuracy, false-alarm rate, and F1, each with a 95% Wilson interval. At the default line it catches 67% of fraud while fast-tracking 44% of claims at a 0.4% fraud rate.

The lesson kept surfacing: for fully structured records, a transparent rule scored against a real answer key beats a confident model that cannot even agree with itself.