sriram1983007 commited on
Commit
c6b1aec
·
verified ·
1 Parent(s): a5b7dc4

Card: real Ethereum payments (300 of 300) and the look-alike probe

Browse files
Files changed (1) hide show
  1. README.md +8 -0
README.md CHANGED
@@ -206,6 +206,14 @@ it did not reject (imitating only 3 characters at each end) were held as new pay
206
  its on-chain poisoning transfer, not by the library's own rule; author-measured. [Dataset, method and every
207
  case](https://huggingface.co/datasets/sriram1983007/sra-real-world-poisoning)
208
 
 
 
 
 
 
 
 
 
209
  **This page is v2 (2026-11-13).** The merged weights here, the [GGUF builds](https://huggingface.co/sriram1983007/SRA-RiskGate-4B-GGUF) and the [Q4_K_M file](https://huggingface.co/sriram1983007/SRA-RiskGate-4B-Q4_K_M-GGUF) are v2: SRA score 0.913 → **0.960**, dispute accuracy 77.8% → **90.2%**, wrongful denials 26.0% → **2.1%**, unsafe payment approvals 0.47% → **0.35%** ([v2 results](https://huggingface.co/datasets/sriram1983007/sra-bench-results)). v1 stays available: `from_pretrained("sriram1983007/SRA-RiskGate-4B", revision="v1")`. The Ollama registry model is still v1 until it is updated. For disputes, pass the computed window in the prompt: `dispute_messages()` in `sra-riskgate` 1.3.0+ builds the exact prompt.
210
 
211
  ### 📚 Paper and articles
 
206
  its on-chain poisoning transfer, not by the library's own rule; author-measured. [Dataset, method and every
207
  case](https://huggingface.co/datasets/sriram1983007/sra-real-world-poisoning)
208
 
209
+ **Real Ethereum payments, and what the model does not do:** on 300 real USDC payments on Ethereum
210
+ mainnet (SRA-Real v1's policy, 1,000 USDC limit), v2 matched the rule label on **300 of 300**:
211
+ 0 of 191 under-limit payments blocked, 0 of 109 over-limit
212
+ approved. Given the real payee in a field it was never trained on, it stopped only **2 of 58** real
213
+ address-poisoning attacks under the limit: look-alike payees are caught by the companion library (179 of 179 with
214
+ `lookalike_check="strict"`), not by the model. Run both. Author-measured; [method and every
215
+ answer](https://huggingface.co/datasets/sriram1983007/sra-real-world-poisoning)
216
+
217
  **This page is v2 (2026-11-13).** The merged weights here, the [GGUF builds](https://huggingface.co/sriram1983007/SRA-RiskGate-4B-GGUF) and the [Q4_K_M file](https://huggingface.co/sriram1983007/SRA-RiskGate-4B-Q4_K_M-GGUF) are v2: SRA score 0.913 → **0.960**, dispute accuracy 77.8% → **90.2%**, wrongful denials 26.0% → **2.1%**, unsafe payment approvals 0.47% → **0.35%** ([v2 results](https://huggingface.co/datasets/sriram1983007/sra-bench-results)). v1 stays available: `from_pretrained("sriram1983007/SRA-RiskGate-4B", revision="v1")`. The Ollama registry model is still v1 until it is updated. For disputes, pass the computed window in the prompt: `dispute_messages()` in `sra-riskgate` 1.3.0+ builds the exact prompt.
218
 
219
  ### 📚 Paper and articles