Core Technology
A decision model built for anti-fraud
The JianCe anti-fraud decision model — a decision engine retrained for anti-fraud: it decomposes judgment into structured questions, solved in parallel, returning probabilities and dispositions in milliseconds. Not a general chat model.
Model positioning
A domain decision model retrained for anti-fraud
Base architecture
Domain retraining on the Qwen3.5 architecture family — parameter size and quantization chosen per scenario, from cloud batch runs to on-device line-by-line inference.
Model form
A decision model, not a generative one: one call answers 7–9 structured questions in parallel, returning probability distributions and confidence.
Question types
noul (yes/no probability), choice (classification), score (0–3 rating) — assembled into question sets per scenario.
Three-layer fusion
Rules/reputation, decision model, and intelligence/graph layers fused: rules for precision, the model for the ambiguous band, intelligence for the last mile.
Multilingual capability scope
The model base covers 201 languages and dialects (Qwen3.5 architecture) — the foundation for cross-language and cross-border fraud content adjudication.
Our fine-tuned decision model is trained on Chinese corpora; performance in other languages is bound by the training and validation scope — no claim of equivalent performance on unverified languages.
Question sets
Not just a score — a reason
Every score traces back to its questions, hit signals, and evidence — supporting customer-service explanation, appeals, and human review.
Returns 0–1 probability. E.g.: is this group black/gray market? Does the SMS ask the user to click?
Returns a class and confidence. E.g.: business domain (12 classes), fraud type.
Returns a 0–3 rating. E.g.: monitoring and disposition priority.
TG seven questions
Real batch run- Main judgment: black/gray or fraud group
- Money channel: money flow / exchange / laundering
- Gambling platform
- Underground services
- Business domain: 12-class choice
- Monitoring priority: 0–3 score
- Needs human review
SMS nine questions
Pilot validation- Main judgment: fraud SMS or not
- Asks to click / asks to transfer / asks for credentials
- Impersonation / fraud type
- Induced action / urgency
- Needs human review
Adjudication output example
Training & fine-tuning
From corpora to a servable decision model
Click any stage for details; play to watch stages advance, checklists tick, and feedback loops flow. Only pipeline stages backed by verified work are shown.
- Corpus collection
- Cleaning & dedup
- De-identification
- Channel & type labeling
- Question-set design
- Labeling spec
- Gold-set split
- Adversarial samples
- Full-parameter CE
- RLCD controlled comparison
- Calibration split
- Recall @ fixed FPR
- Calibration metrics
- Red-team variants
- Throughput & memory
- Staged rollout
- Shadow mode
- Monitoring & rollback
- Drift review
Data preparation
A multi-source corpus pool of authorized de-identified samples, public data, synthetic/red-team samples, and customer-owned data. TG: 6.6M Chinese group names and 350M in-group messages; voice: 28,511 voice-text pairs over 307 hours.
Data flywheel
Every judgment trains the next one
The model is not a static product but a flywheel driven by real data, rule feedback, and human review.
IM corpora
Real batch run6.6M group names · 350M messages
Voice corpora
Training in progress28,511 voice-text pairs · 307 hours
Web / app
Training in progress425 HTML files & screenshots · 196 reviewed
Carrier-side authorized data
Real batch runAuthorized domains / certificates / traffic / fraud results
Performance benchmarks
118.5ms and 69 minutes
Turning bulk judgment into operational capacity: TG figures are from real batch runs; SMS and device-side figures are demo-environment metrics.
Fraud probability distribution (real run)
Bimodal: <0.3 and ≥0.9 hold 87% of mass; the ambiguous band is only ~13% — human review volume stays controllable.
Deposit-system profile (real run)
10,304 deposit-based public groups: P50 ¥6,000, P90 ¥50,000, max ¥6M — deposits are access fees, forming an independent suspicious-system profile.
Fusion & hard negatives
Strict by default — never at the cost of official services
A model alone would kill legitimate verification codes; rules alone would miss variants. Fusion requires gating and hard negatives.
if main judgment ≥ 0.3 or rule hit → trigger sub-question boostingscore = min(1, main + 0.25×supporting signals + 0.10×impersonation)final = max(score, rule score, strong-link score)Official alert whitelist (12381 / 96110 / National Anti-Fraud Center) passes directlyHard negatives
Small-sample trials (synthetic)
Deployment forms
Cloud, on-premise, device, gateway
Small-parameter and quantized versions let the model go where large models cannot.
Cloud API
Real batch run- SaaS / batch runs
On-premise
Capability ready- Carriers / banks / police
On-device inference
Pilot validation- Call / behavior alerts
Gateway embedding
Pilot validation- SMS / traffic entry