BackOphirAutomation
LIVE DEMO · ILLUSTRATIVE

Backbone module · Wave 3 & annual run

The validation your model must pass every single year.

The Baseline Standards are reported to require governance over AI and machine-learning models, with independent validation at least annually — covering accuracy, drift and bias. (Wording paraphrased from reporting on the circular — verify verbatim against the original CBN PDF.) Independent is the operative word: your detection vendor cannot mark its own homework. Ophir validates the model — we never build, sell or replace it.

Mockup 1

Scorecard vs regulatory thresholds

0.0%
✓ PASS
Accuracy
threshold ≥ 90% · +4.1 pts headroom
0.0%
✓ PASS
Precision
threshold ≥ 85% · +3.7 pts headroom
0.0%
✓ PASS
Recall
threshold ≥ 88% · +3.3 pts headroom
0.0%
✓ PASS
False-positive rate
threshold ≤ 6% · 1.8 pts inside limit

Thresholds are the bank's model-risk appetite as approved by its model governance committee · illustrative figures.

Mockup 2

12-month drift monitor

Population-stability score by month

Trailing 12 months · lower is more stable · sample data

0.100.25monitor ≥ 0.10action ≥ 0.250.00AugSepOctNovDecJanFebMarAprMayJunJulDrift detected · PSI 0.19Retrained + revalidated
within tolerance band drift event — Feb retrained + revalidated — Mar

The February event is the point of the annual run: drift was caught, the vendor retrained, and Ophir revalidated before the next cycle — all evidenced.

Mockup 3

Bias & fairness by segment

Detection recall by segment

vs portfolio recall 91.3% · tolerance ± 2.0 pts · sample data

Individual
retail customers
Recall
91.0%
Deviation
−0.3 pts
✓ within tolerance
SME
small & medium enterprises
Recall
91.8%
Deviation
+0.5 pts
✓ within tolerance
Corporate
corporate & institutional
Recall
90.6%
Deviation
−0.7 pts
✓ within tolerance
Region — North
branch & agent network
Recall
89.9%
Deviation
−1.4 pts
◐ monitor
Region — South
branch & agent network
Recall
91.9%
Deviation
+0.6 pts
✓ within tolerance

Region — North sits inside the ±2.0-pt tolerance band but within 0.6 pts of its edge, so it is held on quarterly monitoring rather than remediation. Independent judgement, documented — not a rubber stamp.

Mockup 4 · interactive

Annual validation report

Compile the examiner-ready pack

Sampling → back-testing → drift → bias → sign-off · ~2 seconds · sample data

Validation run

Sampling — 12 months of alerts and confirmed outcomes drawn
Back-testing — model output replayed against investigator dispositions
Drift analysis — population stability checked month by month
Bias review — recall compared across segments and regions
Sign-off — independent validation memorandum sealed

The sealed, examiner-ready validation report appears here after the run.