Pix Race
Compare Jev scam judgments and deterministic money-flow tracing on synthetic Pix messages.
Added to Jevfast
How it uses Jev
Pix Race sends each simulated Pix message, with its sender, to Jev for four questions in one System One request: scam likelihood, message type, urgency, and whether it asks for a Pix payment. Rust applies its scam decision threshold. If the message is classified as a scam, it uses the Pix key to trace the synthetic transaction graph. A Python and DeepSeek side processes the same messages for comparison.
What Jev decides
Illustrative synthetic Pix message classification request shapeExample answers · not a recorded Jev response · Source ↗
Question 1 · golpe
YOUR APP
INSTRUCTION
Decide whether the message is a scam or fraud attempt, such as impersonation, a fake bill, fake bank alert, fake prize, or pressure to pay.
STATE
Source-defined request state contains the message sender and message text. The repository describes its messages and transaction graph as synthetic; this chart does not reproduce a specific message or runtime answer.
JEV · NOUL
YesNo
Question 2 · tipo
YOUR APP
INSTRUCTION
Choose the message category from the source-defined classification labels.
STATE
Source-defined request state contains the message sender and message text. The repository describes its messages and transaction graph as synthetic; this chart does not reproduce a specific message or runtime answer.
JEV · CHOICE
- troca_numero
- boleto_falso
- falsa_central
- premio
- pix_errado
- cobranca_legitima
- pessoal
- outro
Question 3 · pede pix
YOUR APP
INSTRUCTION
Decide whether the message asks the reader to make a Pix transfer or payment.
STATE
Source-defined request state contains the message sender and message text. The repository describes its messages and transaction graph as synthetic; this chart does not reproduce a specific message or runtime answer.
JEV · NOUL
YesNo
App workflow
- Load the synthetic race data
The Rust server loads the messages, Pix-key map, reference scam metadata, and generated transaction graph.
- Send each message to Jev
The Rust client submits the sender and text with all four typed questions to the TypeSafe System One endpoint. The server can process messages serially or in a bounded burst.
- Classify the result
The client parses the Noul, Choice, Score, and token-usage fields. The server converts scam probability to a scam, review, or OK verdict and broadcasts the decision and timing data to the browser.
- Trace scam payments
For scam verdicts, the server maps the Pix key to a graph node and runs the graph trace from the message timestamp.
- Compare both sides
The browser displays Rust plus Jev beside Python plus DeepSeek and reports their results against the synthetic reference labels.
Why it is interesting
The demo separates model judgment from follow-on graph analysis. Jev classifies the message, then deterministic Rust code traces the payment path only for a scam verdict. The split-screen race compares Jev and DeepSeek on the same generated messages and graph data.