SRA-Real v1: payment checks on real Base USDC payments

227 test cases built from real USDC payments settled on Base mainnet (EIP-3009 transferWithAuthorization / receiveWithAuthorization, 1–2 November 2026). It is the real-payment companion to the synthetic sra-stablecoin-risk-bench: same row format, same prompts, same scorer (score.py), so any model evaluated there can be run here unchanged.

Kind of case	Cases	What is real	What was changed
Clean	151	Payer, payee, amount, validity window, nonce and signature	Nothing (a neutral memo is added)
Not yet valid	10	Everything, including the payment's own future validAfter	The check time is set before validAfter
Expired	11	Everything	The check time is moved past validBefore
Over the limit	11	Everything	The policy limit is set to half the real amount
Payee sanctioned / payer sanctioned	11 / 11	Everything	The sanctions screen reports a hit
Bad signature	11	Everything	The signature check reports a failure
Prompt injection	11	Everything	Instruction text is added to the memo

Every row records its source transaction hash, block, and exactly what was changed (source.change).

Results (measured by the author, 2026-11-12)

	SRA-RiskGate-4B v1	SRA-RiskGate-4B v2	Base Qwen3-4B-Instruct-2507
Decision accuracy	100% (227/227)	100%	78.4%
Real payments approved (151 clean)	151/151	151/151	137/151
Unsafe approvals	0%	0%	28.9%
Reason-code (flag) F1	0.994	1.000	0.017

Predictions, scores and run settings for every run are in sra-bench-results/sra-real-v1. The base model's most common error on real traffic: it read the raw on-chain value (e.g. 1000) as dollars and called the correctly decoded 0.001 USDC an "amount mismatch". Real agent payments are tiny: 143 of the 151 clean payments are under 0.10 USDC (median 0.01 USDC, largest 273.75 USDC).

How to read the perfect scores. The 76 changed cases use the same kinds of change as the synthetic benchmark, and the injection texts are in its style, so they show that the model's checks carry over to real payment data (real amounts, real time windows, real signatures), not that it resists new kinds of attack. The strongest real-world result is the clean set: every real payment approved, none blocked.

How it was built

1. A 24-hour sample of Base mainnet USDC settlements (1,500 transactions; 1,049 direct EIP-3009 calls).
2. Each payment's calldata was re-read from the chain and the payer's EIP-712 signature verified by recovering the signer (domain "USD Coin" v2, chain 8453, USDC 0x8335…2913). 2 payments whose signature does not recover to the payer (smart-contract wallets, ERC-1271) were excluded.
3. Only payments valid at settlement were kept. At most 25 payments per payee, so one busy merchant cannot dominate: the 227 cases come from 69 payees and 165 payers. Agent payments on Base are concentrated in a few merchants; the cap skipped 820 payments.
4. Two thirds became clean cases, one third got exactly one documented change.
5. Labels follow the benchmark's own conventions (decision, risk level, flags, required actions), and every label was checked to score perfectly against itself with score.py --baseline oracle.

build_sra_real.py is the exact script used.

Policy, the same for every case: limit 1,000 USDC (except the over-limit cases), Base USDC allowed, no risk attestation required. Real payments carry no attestation.

Privacy

Payer and payee addresses are replaced by consistent stand-ins: the same real address always maps to the same stand-in, and the mapping is kept private. The stand-ins keep wallet addresses out of the text that models see and that people quote. They are not anonymisation: each row keeps its transaction hash, and on-chain data is public, so anyone can look up the real payment. The signature in each row is the real one; it verifies against the real addresses (look up the transaction), not against the stand-ins. The sanctions hits in the changed cases are synthetic and say nothing about the real addresses.

Files

File	Contents
data/benchmark.jsonl	The 227 cases in the benchmark format (id, task, format, perturbations, input, label, source)
data/sft.jsonl	The same cases as the model's exact prompts, aligned by line
label_sheet.csv	100 cases for blind human labelling (no answers in it)
build_sra_real.py	The build script

Run a model on it

hf download sriram1983007/sra-real --repo-type dataset --local-dir sra-real
hf download sriram1983007/sra-bench-results run_eval.py --repo-type dataset --local-dir .
hf download sriram1983007/sra-stablecoin-risk-bench score.py --repo-type dataset --local-dir .
python run_eval.py --name my-model --backend openai --base-url http://localhost:8000/v1 --api-key-env KEY \
    --model my-model --bench-file sra-real/data/benchmark.jsonl --sft-file sra-real/data/sft.jsonl

Limitations

• Small: 227 cases from one day on one chain, one token (USDC) and one payment format (EIP-3009).
• Labels come from the benchmark's rules, not human judgement. label_sheet.csv is there so independent labellers can check them.
• No disputes: real dispute data is not public on chain.
• Measured by the author; no third party has reproduced these results yet.

Related

• Model: SRA-RiskGate-4B
• Synthetic benchmark: sra-stablecoin-risk-bench
• All results: sra-bench-results
• Code: GitHub