Every motor claim is sorted into one of four handling categories by a fixed, published rule, then the fraud flags are checked against the real outcome. Same claim, same category, every time.
Decision support only. A human assessor owns every flag and every repudiation. No claim is ever auto-denied.
Classified 0 real claims with a deterministic yardstick (no model, no API key). The accuracy below is measured against the real fraud outcome on each claim.
Full population · 0 true fraud rate
Where every claim landed. Each claim sits in exactly one category, decided by a fixed point score over its facts. Green clears with minimal effort, red is a denial candidate. The split is fixed; it does not change between runs.
0%
fast-tracked, ready to clear with minimal effort
0%
routed to a human (approve, investigate, or repudiate)
0%
flagged for fraud (investigate or repudiate)
The four handling categories
Tap a category to list those claims below.
Show:
Pick any claim and see how the rule classifies it: the points it scored, the category, and whether it was actually fraud. This is the same rule the whole batch uses, so a claim's category here always matches the triage and the scorecard.
The rule was built from the claim facts only; the real fraud outcome was held back and used only here, to score it. These are full-population results over every claim, so the confidence ranges are tight.
Can you trust the fraud flags?
The two numbers a fraud team cares about most, at the current review threshold.
What actually happened
Every claim falls into one of four boxes once the flags are compared to the truth. Updates when you move the review threshold.
Truly fraudulent
Truly legitimate
Flagged
Caught fraud
0
flagged and it really was fraud
False alarm
0
flagged an honest claim
Cleared
Missed fraud
0
cleared a claim that was fraud
Correctly cleared
0
cleared an honest claim
The full numbers
Each with a 95% confidence range (very tight here, because every claim is scored).
Does a higher score mean more fraud?
True fraud rate at each risk-point level. If it climbs left to right, the score is doing its job.
Fraud rate by accident area
A simple sense-check of where real fraud concentrates.
The four categories are fixed. This dial is a separate question for the fraud team: at what risk-point level should a claim be flagged for review? Default is 4 and above (Investigate or Repudiate). Move it and watch the trade-off in real numbers.
Flag a claim for review when risk points are at least
4
1 · flag almost everything8 · flag only the clearest
Fraud caught of all real fraud
0%
Flags that were right when flagged, how often correct
0%
Honest customers bothered false alarms per 100 honest
0%
Claims sent to review your team's workload
0
The trade-off, drawn out
Each point is a possible threshold. Up and to the left is better. The marker is where the dial sits now.
The four categories
How the whole book splits across the fixed categories.
An honest account of how a claim is classified, why it is the same every time, and what is and is not measured.
The yardstick
A claim scores points from the facts the data proves actually predict fraud. The total decides the category. Nothing else.
Signal in the claim
Points
Policy holder at fault
+2
All Perils policy (Collision +1, Liability +0)
+2
Recent address change at claim (under 6 months or 2 to 3 years)
+2
Accident at policy start (zero days policy-to-accident)
+2
Rural accident
+1
Vehicle price at an extreme (under 20k or over 69k)
+1
Vehicle 0 to 4 years old
+1
Total points
Category
Action
0 to 2
Fast track
auto-clear, minimal effort
3
Approve
pay after standard processing
4 to 5
Investigate
refer to the fraud unit before any decision
6 or more
Repudiate
recommend denial, a person decides
Why Results Stay Consistent
1
The category is arithmetic, not a guess
The points are a fixed sum over fixed fields. The same claim always scores the same total and lands in the same category, whether it is reviewed alone or inside the whole book.
2
No model in the classification path
There is no language model, no temperature, no randomness deciding the category. The rule is published above and you can check any claim against it by hand.
3
The truth is held back, then used to score
The fraud label is never an input to the rule. It is revealed only on the scorecard, to measure how often the flags are right.
What is measured and what is not
Fraud is scored. The data carries a real fraud label, so the flags (Investigate and Repudiate) are measured against it. That is the scorecard.
The split into four categories is operational routing. Built from the facts, grounded in the signals that predict fraud in this data, and applied identically to every claim.
Repudiate means strong fraud grounds here. This public dataset has no coverage or policy-status field, so Repudiate is the top fraud tier. A real system would also fire it on coverage checks.
The rules that never bend
Decision support only. A human assessor owns every flag and every repudiation.
No auto-denial of claims. A flag is a trigger to review, not an accusation.
The yardstick is fixed and published, so results reproduce exactly.
Public data, used for portfolio demonstration only.