KERNEL ONLINE ·
FRAUD DETECTION ENGINE

Stop fraud without
breaking good payments.

SignalBrain-OS combines account, device, payment, graph, and policy signals into signed fraud decisions. It catches attacks, explains the evidence, and creates case packets that analysts can act on.

EVIDENCE UPDATE · JULY 18, 2026

Retail-app and tap-to-pay fraud strategy, replayed A/B.

Titan v2 combines account-recovery timing, device age and trust, wallet-token provenance, gift-card concentration, cross-store velocity, and decline probing. The same fixed 27-attempt synthetic cohort was replayed end-to-end through the shared Sentinel cross-domain runtime.

83.33%Titan v2 fraud recall, up from 33.33%
+$940Additional simulated fraud value prevented
4.76%Legitimate friction, up from 0%
83.33%Flag precision, down from 100%

Success: four additional fraud patterns were stopped. Failure: one mature stolen wallet token remained undetected, and one legitimate post-recovery purchase was challenged. These results prove deterministic test behavior—not production efficacy or issuer/processor validation.

Read the method, failures, and reproducible proof →
WHAT THE EVIDENCE PROVES

Integrity proof and decision accuracy are different claims.

Portable decision evidence

SQL Server Ledger and similar database features provide meaningful cryptographic tamper evidence inside a centrally administered database, including verification against published digests. If one database operator controls the complete workflow, that may satisfy much of the integrity requirement.

Titan's distinction is portable, signed decision evidence that can be checked outside the originating database across agent, organization, and system boundaries. Titan currently implements signed, tamper-evident receipts with offline public-key verification. Independent multi-party key custody and external anchoring remain operational validation work, not completed claims.

Auditable policy, not proven prediction

Titan's fraud path is not presented as a conventional high-parameter model trained solely on millions of historical transactions. The implemented path applies constrained, reviewable policy to structured transaction, device, wallet, velocity, evidence-freshness, and runtime-health inputs.

Deterministic repeatability does not establish predictive efficacy. Production claims require representative labeled outcomes measuring fraud recall, precision, false-positive or legitimate-friction rate, review outcomes, latency, and receipt integrity.

FRAUD TYPOLOGIES

Built for the fraud your team actually sees.

Fraud is not AML. It is account compromise, payment abuse, identity manipulation, and loss prevention under conversion pressure.

ATO

Account Takeover

Detect new device, impossible travel, credential reset, phone change, and first high-value transaction in one timeline.

Testing

Card Testing

Identify rapid low-value attempts, issuer decline patterns, merchant concentration, and linked IP/device reuse.

Abuse

Refund & Promo Abuse

Score refund velocity, repeat merchant disputes, policy exploitation, multi-account couponing, and suspicious fulfillment loops.

Networks

Mule Accounts

Trace funds through new beneficiaries, rapid cash-out, shared devices, shared addresses, and clustered counterparty risk.

Identity

Synthetic Identity

Combine identity proofing, account age, behavioral inconsistency, phone/email risk, and payment instrument reuse.

Scams

Authorized Push Payment Risk

Flag urgent narratives, new payees, unusual amounts, remote-access indicators, and sudden behavior change before transfer.

GRAPH RISK

Find fraud rings, not just bad transactions.

Shared-entity graph

SignalBrain links accounts, cards, devices, IPs, addresses, beneficiaries, merchants, wallets, phone numbers, and payment instruments. A single transaction can inherit risk from its neighborhood.

Device FanoutOne device touches 18 accounts in 24 hours.
Beneficiary ReuseNew payee appears across unrelated customers.
IP ClusteringCheckout bursts share VPN/proxy infrastructure.
Payment InstrumentCard or wallet token recurs across accounts.

Evidence receipt

Every fraud decision returns a compact proof packet: risk score, reason codes, linked entities, model version, policy version, and Merkle proof. It is built for analyst review, dispute handling, and control testing.

{
  "decision": "HOLD",
  "risk_score": 0.86,
  "reason_codes": [
    "DEVICE_FANOUT",
    "NEW_BENEFICIARY",
    "ATO_TIMING"
  ],
  "linked_entities": 7,
  "case_packet": "case_31a...",
  "certificate_id": "cert_b92...",
  "merkle_proof": "proof_6db..."
}
OPERATIONS

From alert to case packet.

01
AlertModel, rules, graph, or external signal escalates the event.
02
AssembleTimeline, account changes, payment history, and linked entities are gathered.
03
DecidePolicy emits approve, step-up, hold, decline, or manual review.
04
ProveReceipt and case packet are signed for audit and analyst handoff.
API EXAMPLE

Fraud evaluation with reason codes.

{
  "domain": "fraud",
  "event": {
    "type": "payment_attempt",
    "customer_id": "cus_8847",
    "amount": 1250.00,
    "currency": "USD",
    "merchant_id": "m_48291",
    "device_id": "dev_91a",
    "beneficiary_id": "ben_42"
  },
  "signals": {
    "password_reset_minutes_ago": 18,
    "new_device": true,
    "new_beneficiary": true,
    "device_account_count_24h": 18,
    "velocity_1h": 5
  }
}
{
  "decision": "HOLD",
  "recommended_action": "MANUAL_REVIEW",
  "risk_score": 0.86,
  "reason_codes": [
    "RECENT_CREDENTIAL_RESET",
    "NEW_DEVICE",
    "DEVICE_FANOUT",
    "NEW_BENEFICIARY"
  ],
  "review_sla": "15m",
  "case_packet": "case_31a..."
}

Fraud decisions with proof, not guesswork.

Deploy beside your payment processor, wallet, identity provider, or case-management stack.

Mobile Payments Cyber Response Request Fraud Review