MODEL VALIDATION / ANALYST TRIAGE DEMONSTRATION

Fraud risk control review

Full row-level event queries, fixed holdout evidence, and constrained review capacity.

Coverage: Jan 2019 to Dec 2020
Holdout: 555,719 events
Review date: 18 Aug 2026

RELEASE DISPOSITIONRETAIN BASELINE

Measured 1% analyst-review queue only. Not production-ready.

Recall at 1%: 51.3%+5.9 pp vs challenger
PR-AUC: 0.160challenger 0.112
Brier: 0.004challenger 0.077
Review slots: 5,557chronological holdout

CONTROL 01 / PUBLIC EVENT QUERY

Complete simulated event register

Query every allowlisted event row. Results are read-only, ordered by event time, and limited to 100 rows per request.

1,852,394 public event rows25 returnedidentity-like source columns excluded

Event time, UTC	Event ID	Merchant / category	Amount	Fraud label	Source
Jan 1, 2019, 12:18 AM	`26`	fraud_Kunde-Sanfordmisc_net	$2.55	NO	fraudTrain
Jan 1, 2019, 12:20 AM	`27`	fraud_Deckow-O'Connergrocery_pos	$64.09	NO	fraudTrain
Jan 1, 2019, 12:21 AM	`28`	fraud_Bruen-Yostmisc_pos	$6.85	NO	fraudTrain
Jan 1, 2019, 12:22 AM	`29`	fraud_Kunze Incgrocery_pos	$90.22	NO	fraudTrain
Jan 1, 2019, 12:22 AM	`30`	fraud_Nitzsche, Kessler and Wolffshopping_pos	$4.02	NO	fraudTrain
Jan 1, 2019, 12:22 AM	`31`	fraud_Kihn, Abernathy and Douglasshopping_net	$3.66	NO	fraudTrain
Jan 1, 2019, 12:22 AM	`32`	fraud_Heller, Gutmann and Zieme

CONTROL 02 / CAPACITY

Queue planning scenario

Change the planning rate to see workload. Recall remains visible only for the measured 1% policy because no other threshold sweep is available.

Holdout review rate0.5%1%2%

Planned review slots5,55751.3% measured recall on the chronological holdout

EVIDENCE 03 / MODEL COMPARISON

Baseline outperforms the challenger at the measured queue constraint

Both policies train on the earlier simulated partition and evaluate on the later one. The challenger only adds class weighting to the same logistic model family. It loses on every reported measure and is not promoted.

Measure

BaselineLogistic regression

ChallengerClass-weighted logistic regression

PR-AUCRare-outcome ranking; higher is better.

16%11.2%

Recall at 1% reviewFraud labels surfaced within fixed capacity; higher is better.

51.3%45.4%

Brier scoreProbability error; lower is better.

0.0040.077

Calibration

Predicted bands vs observed fraud rate

Baseline only

CONTROL 04 / DATA BOUNDARY

Local processing and public evidence are separate zones

The complete simulated source is published through allowlisted analytical fields. Identity-like raw columns and decisioning capabilities remain outside the public view.

1. Simulated source1,852,394 events validated against the approved source contract.
2. Governed pipelineValidation, point-in-time features, and chronological evaluation run locally.
3. Public event viewEvery event row through seven allowlisted analytical fields.
4. Public registerBounded queries, monitoring, and precomputed evaluation evidence.

Not available here: identity-like raw columns, personal identifiers, model scores, scoring endpoints, payment decisions, or payment-processing behavior.

EVIDENCE 05 / COVERAGE LEDGER

Complete event publication by source partition

These source totals reconcile the public event view to all 1,852,394 ingested rows.

Source partitionEventsLabelled fraud ratePeriod

fraudTest555,7190.4%Jun 2020 to Dec 2020

fraudTrain1,296,6750.6%Jan 2019 to Jun 2020

LOCAL SOURCE TOTAL1,852,3949,651 labelled fraud events

The validation register is opening the read-only event view.