The fraud result, the conversion cost, and the miss—all in one receipt.
We submitted the identical deterministic cohort of 27 synthetic retail-app and tap-to-pay attempts over HTTP to /v1/sentinel/evaluate. Six were labeled fraud and 21 legitimate. Both policies traversed domain routing, legacy payment safety, fast council, domain enforcement, and dry-run certification.
Shared-runtime A/B
| Metric | Baseline v1 | Titan v2 | Delta |
|---|---|---|---|
| Fraud recall | 33.33% | 83.33% | +50.00 pp |
| Simulated fraud value prevented | $2,150 | $3,090 | +$940 |
| Legitimate friction | 0.00% | 4.76% | +4.76 pp |
| Flag precision | 100.00% | 83.33% | −16.67 pp |
What Titan added
Failure evidence
Success: three additional labeled fraud attempts were stopped after the shared runtime's existing safety layer strengthened the baseline.
False negative: one mature stolen wallet token with otherwise normal telemetry was approved.
Legitimate friction: one valid purchase shortly after account recovery was escalated to phishing-resistant step-up.
Mature-token takeover: designed next, not claimed now
A mature stolen token can keep ordinary device, velocity, merchant, and spending telemetry looking normal. The correct response is a targeted chain-of-custody layer—not blanket friction.
Provenance: record issuance assurance, recovery path, device enrollment, identity tier, and any SIM-swap or low-assurance link history.
Ownership continuity: compare the present controller with historical device, IP, ASN, carrier, geo, and app fingerprints; enforce token→account→canonical-device binding.
Adaptive policy: route only anomalous mature tokens to independent proof of control for high-risk purchases, and learn from privacy-reviewed issuer decisions and delayed chargebacks.
Status: this layer is designed but not represented in the A/B result above. Production enforcement requires wallet-provider/issuer inputs and a focused hard-positive/hard-negative calibration cohort.
Inspect and reproduce
PYTHONPATH=src:. python3 scripts/simulate_retail_payment_fraud_ab.py
Evidence boundary: this is deterministic synthetic runtime evidence, not processor, issuer, wallet-provider, chargeback, or production loss validation. Thresholds require privacy-reviewed production calibration before enforcement.