Demo #13 · Fraud · Instant Payments · Processors Hub

6-signal score model Sub-200ms latency budget @ainumbers.co/instant-pay-fraud-v1

Decide in 200 milliseconds. Live with the call for 10 years.

Instant-payment rails are irrevocable — once the money lands, the fraud loss is final. The decision must happen in the sub-second window between auth and send. This lab models a six-signal fraud score against an adjustable threshold and intervention budget, surfaces the resulting confusion matrix and expected dollar loss, and exports the chosen policy as an AP2 mandate the runtime enforces and the fraud-ops team can audit.

Zero PII · Client-side Synthetic 100K-tx population · illustrative, not predictive Last Reviewed · 2026-05-13

FedNow · RTP · SEPA Instant — irreversibility means real-time matters

FedNow (Federal Reserve, July 2023) and TCH RTP (Real-Time Payments, 2017; $1M cap from February 2024) are credit-push, irrevocable rails. Once accepted by the receiving institution, recall is at the receiver's discretion. SEPA Instant (EU; SCT Inst mandatory for euro-area banks from Jan 2025 under Regulation (EU) 2024/886) is the European equivalent. Reg E protections for unauthorised transfers cover ACH; for consumer-initiated push payments on RTP/FedNow, recovery depends on the receiving bank's cooperation — which is rarely guaranteed.

Sources: Federal Reserve FedNow service guide · TCH RTP rulebook (Feb 2024 cap) · EU 2024/886 (Instant Payments Regulation) · 12 CFR §1005 (Reg E)

§1 · Fraud Signals 6 / 6 active

Score model — six signals

Each signal contributes a 0–100 score, weighted into a composite. The illustrative weights below are typical; calibration is the heart of the work. Toggle a signal off to model fallback behaviour when a feed is unavailable.

§2 · Decision Policy Threshold · intervention

Threshold & intervention budget

Below threshold: pass. At threshold: step-up auth (selfie / device-bind / call-back) inside the intervention budget. Above ceiling: hard refuse with refer.

Decision threshold (composite score) Tx scoring at-or-above this triggers step-up auth.

Hard-refuse ceiling Tx at-or-above this score are refused outright, no step-up offered.

Intervention budget (% of tx allowed step-up) Step-up adds friction; budget caps daily user-experience cost.

2.5%

§3 · Population & Loss Model Synthetic, illustrative

Calibration inputs

Daily transaction volume Programme-level instant-payment volume across all customers.

100,000

Base fraud rate (per 10,000 tx) Typical instant-pay range is 5–25 per 10K for unprotected programmes.

Average fraud loss per missed tx (USD) RTP cap $1M; consumer-fraud avg ~$1,500; BEC avg ~$30K.

$1,800

Recovery rate (% of missed-fraud losses recovered) RTP/FedNow recovery via beneficiary-bank cooperation is typically < 30%.

15%

§4 · Rail Scope FedNow · RTP · SEPA Inst

Which rails the mandate covers

The mandate identifies which instant-payment rails the policy binds. Different rails have different irrevocability and recovery profiles.

§5 · Output Live confusion + AP2 mandate

Decision economics

Precision

—

TP / (TP + FP)

Recall

—

TP / (TP + FN)

Step-up rate

—

vs intervention budget

Expected $ loss / day

—

missed × avg × (1−recovery)

Confusion matrix (per day)

Predicted fraud

Predicted clean

Actual fraud

—

true positive

—

missed (loss)

Actual clean

—

false positive

—

true negative

Latency budget · 200ms — composite + step-up handoff estimate

AP2 v1.0 · valid · @ainumbers.co/instant-pay-fraud-v1

Compose with

Different from #11 — AML TM Rule Builder (BSA/AML, post-hoc detection, SAR-focused); this demo is real-time fraud-loss prevention with a hard latency budget. Pair with #23 — AP2 AML Mandate Builder to bind both fraud and AML policy into the same runtime. Position in the 259-tool atlas: Tool Chain Composer, T04 + T26 in the Fraud & Risk cluster.