Chief Risk Officer · dispute policy

Set the policy. See what it costs before it ships.

Autonomy is a policy decision, not a model setting. Move a parameter and the engine re-runs over the demo book and a labelled history, showing which cases change hands, how many credits would have gone to invalid claims, and what the score really means.

Policy parameters

POL-B1A78921

Default policy.

Auto-resolve threshold (evidence score)0.95

Autonomous-credit ceiling by segment

Mass$1,000

Affluent$1,000

Private Client$1,000

File floor (below it: credit and write off)$25

Categories the agent may resolve alone

Merchandise / services not receivedFraud - card-not-presentDuplicate processingNot as described / defective

Stored only in this browser. Nothing you change here is seen by anyone else.

Effect on the 38-case demo book

The book is authored to exercise every outcome, so read the direction of change, not the level.

27 of 38

resolved inside policy

same as default

$11,054.30

credited without an analyst

same as default

4.6 h

analyst time (modeled)

same as default

$18

blended cost per dispute (modeled)

same as default

No case changes hands under this policy

Backtest: 1,000 labelled synthetic disputes

synthetic labels

Each closed dispute carries an outcome - was the claim valid? - drawn from a separate, documented process (see Method below). The engine runs over all of them under your policy. This shows the workflow a bank runs on its own history in shadow mode; it is not evidence of accuracy on real disputes.

44.3%

resolved inside policy

same as default

87.8%

of those were valid claims (precision)

same as default

credited, then found invalid ($12,207 exposure)

same as default

238

valid claims still sent to an analyst

same as default

Exposure is the amount credited before any representment or reversal recovers it; it is not a loss figure.

By category

Category	Disputes	Inside policy	Precision	Credited, invalid
Merchandise / services not received	373	225	87.1%	29
Fraud - card-not-present	326	139	84.2%	22
Duplicate processing	111	79	96.2%	3
Not as described / defective	190	0	-	0

Threshold trade-off (other parameters as set)

Dashed line: your threshold. Scores cluster at the top of the range, so thresholds between 0.95 and 0.97 barely change the outcome - the ceilings and category switches often matter more.

Calibration: what a score actually means

If the score were a probability, each bar would reach its marker. It is not: the weights are hand-set, so the model-risk packet reports the calibrated rate, and the policy threshold is chosen on it.

Score bands on the evidence score; n = synthetic disputes in each band.

Method: how the synthetic labels are generated

P(claim valid) = logistic(z), z = -0.2; not received +1.7 without delivery proof, -1.6 with it; card-not-present fraud +1.6 without 3-D Secure, -1.3 with it; duplicate +3.2 if found, -1.2 if not; not as described +0.3; transaction not matched -0.9; confirmed first-party misuse -2.0, two or more prior disputes -0.7, otherwise +0.4; merchant velocity high +0.5, elevated +0.25, low -0.15; minus 0.25 x log10(amount / 100); call-centre channel -0.1; plus Normal(0, 0.5) noise.

These weights differ from the engine's score weights on purpose, and include effects the score ignores (amount, channel, noise). Feature mix: 36% not received, 34% card-not-present fraud, 10% duplicate, 20% not as described; amounts log-normal around $140. Seeded, so every visitor sees the same history.