15 Downloads Updated 1 month ago
ollama run sriram1983007/sra-riskgate
Checks a stablecoin payment before an AI agent sends it: approve, hold or reject, with reason codes.
A 4B model, fine-tuned from Qwen3-4B-Instruct-2507, that checks a stablecoin payment before an AI agent sends it and answers approve, hold, or reject with reason codes.
Build: Q4_K_M GGUF, payment risk-gate system prompt, temperature 0, [8K or 256K] context window.
Intended use
Pre-signing payment checks: one decision per payment, with the amount, payee, and context as input. This is not a general assistant. Do not use it for other tasks.
Results (self-reported)
Pre-sign gate decisions: 99.5% correct on a 1,175-case held-out split of synthetic cases. Base model: 59.8%. Unsafe-approval rate (pre-sign): 0.47%. Base model: 34.6%. Post-settlement dispute adjudication: 77.8%.
Known limitation: dispute adjudication has an over-denial bias. About a quarter of valid refunds are wrongly denied. Do not use this model alone for dispute decisions.
These results come from synthetic test cases and have not been independently reproduced.
Links
Model: https://huggingface.co/sriram1983007/SRA-RiskGate-4B
Benchmark: https://huggingface.co/datasets/sriram1983007/sra-stablecoin-risk-bench