0563GitHub
Jev Guard: triage exchange deposits and withdrawals
Jev judges transfer risk and behavioral patterns; Java rules decide whether the case passes, freezes, or goes to an analyst.
The real TypeSafe client, hard-rule layer, and routing design were inspected on October 1, 2026. Hard rules run before the model; direction-aware code composes the final decision, and provider failures go to manual review. Jev’s suggested route is logged for comparison rather than executed directly. The included mock demo is deterministic. No live exchange deployment, fraud-detection accuracy, or financial outcome is claimed here.
klauswg/jev-guardREADME ↗
# jev-guard
**Real-time risk triage gateway for crypto exchange deposits & withdrawals — Jev (TypeSafe System One) as the triage layer, deterministic code as the judge.**
[中文 README](README.zh-CN.md)
Jev is a decision model that answers structured questions over a state. This project answers one question with it: *given a deposit/withdrawal and its 24h behavioral features, who should handle it — auto-pass, human review, or freeze?* The answers (risk level, behavioral pattern, freeze probability) come from Jev; the **thresholds, the routing matrix, and the final action stay in Java** — because that's where they can be unit-tested, audited, and tuned.
## Why this architecture
- **Jev triages, code decides.** A hard-rule layer (blacklist / sanctioned counterparty / amount cap) vetoes before any model call. The model never gets the last word.
- **High-risk withdrawals freeze, never downgrade.** Once broadcast, a withdrawal is irreversible; a high-risk deposit is already on-chain, so it becomes account review instead. The direction × action matrix encodes this asymmetry.
- **`route` is recorded, never executed.** Jev's own routing suggestion is kept as an observation-only column so we can measure how often "model routes directly" disagrees with "code composes the decision" (72% agreement on our calibration set — that 28% gap is exactly why composition stays in code).
- **Degrade to human review, never to pass.** Upstream timeout/5xx/429 → `MANUAL_REVIEW`. Better to annoy an analyst than to let one through.
## Architecture
```
transfer event
│
▼
FeatureBuilder ── 24h window (Caffeine, in-memory): count/volume/burst/age/rapid-in-out
│
▼
HardRuleEngine ── blacklist / sanctioned / >$500k ──hit──► FREEZE / MANUAL_REVIEW (no model call)
│ miss
▼
Sta