ZWYJ · Anti-fraud decision engine
Anti-fraud judgment, back to milliseconds
The JianCe anti-fraud decision model — a decision engine built for anti-fraud. One call, seven questions in parallel, returning probability, type, and disposition: not a paragraph of generated text, but an explainable, auditable judgment.
Real-time programmatic concept demo · not production data
Validated throughput
Turning bulk judgment into operational output
The value of an decision model is not only accuracy — it is turning full-volume data into actionable, auditable conclusions within acceptable time and cost. Figures below are from real batch runs and demo environments.
Paradigm shift
From token generation to parallel adjudication
Anti-fraud demands high-frequency, low-latency, explainable, reviewable bulk judgment. A generative LLM is not the right shape for it.
General-purpose generative LLM
- Token-by-token generation, thinking before concluding
- Second-level latency — cannot enter real-time paths
- Output is natural language: hard to audit and reconcile
- Full-volume invocation is too costly to operate at scale
JianCe anti-fraud decision model
- One call answers 7–9 structured questions in parallel
- Millisecond latency — fits gateways and devices
- Output is probability + type + priority + review advice
- noul / choice / score question types trace results back to evidence
final = max(main judgment, sub-question boost, rule score, strong-link score)Boosting only when main judgment ≥ 0.3 or a rule hits; official whitelist passes directlySolution overview
One model, three fused layers, reused across the chain
Not another system — an adjudication engine embedded into existing anti-fraud paths. Rules deliver precision, the model covers the ambiguous band, intelligence closes the last mile.
Rules / reputation layer
The high-precision, low-cost first gate
- Port and number-range reputation
- Signature & template registration
- Frequency and behavior patterns
- Blacklisted links and domain age
Decision model layer
The semantic core for the ambiguous band
- noul probability · choice classification · score rating
- 7–9 questions answered in one parallel call
- Probability and confidence outputs
- Two question sets: TG seven questions, SMS nine questions
Intelligence / graph layer
Cross-scenario correlation and closed-loop disposal
- Entity extraction (domains / cards / accounts)
- Cross-scenario entity correlation
- Family clustering and expansion
- Disposition feedback loop
Decision pipeline
Five pipeline stages, one judgment in milliseconds
Click any stage to inspect inputs and outputs; play, pause, or replay the full data stream.
Normalize
Neutralize evasion variants
Full/half-width unification, traditional/simplified conversion, homophone and split-character restoration, zero-width cleanup, short-link expansion. The most common black-market evasion tricks are flattened here.
Input / output · IN
- Raw text
- Short links
- Obfuscated variants
Input / output · OUT
- Standardized text
- Expanded links
Use cases
Wherever fraud reaches, adjudication follows
Five reach scenarios plus three capability outputs, covering the chain from luring to money flow. Each scenario has its own workflow and interactive demo.
Proof wall
Real batch runs, real calls, real demos
Not capabilities on a slide — systems that have been run, clicked, and recorded. Each card states its data provenance.
58,133 groups adjudicated in full
69.2 minutes end to end, 14.0 groups/second, 30,047 black/gray-market groups identified.
99.1% agreement across two full runs
Only 546 disputed items — output is stable and reviewable.
SMS gateway demo: 46,200 messages/day
82ms judgment-to-alert, zero false alerts on official numbers, 1,483 active ports profiled.
On-device training data assets
28,511 voice-text pairs, 307 hours of audio, 12,406 elder-fraud dialogs.
Web detection training set
425 HTML files and screenshots, 196 human-reviewed main tasks, a 14-class label system.
Small-sample live trials
All 11 fraud samples ≥ 0.58 (8 at ≥ 0.91); 2 normal chat samples ≤ 0.04.
Partnership
Pick one scenario; validate offline in 2–4 weeks
No platform needed first. Choose a scenario, validate the question set and judgment quality on authorized de-identified samples, then decide the scaling path.
Pilot
Pick a scenario; complete offline validation and a 1,000-item human evaluation in 2–4 weeks.
Co-development
Turn your data boundary and disposition workflow into a deployable adjudication solution.
Capability integration
Risk scores for domains / IPs / apps / cards via API, on demand.