SRA-RiskGate-4B (GGUF)

GGUF builds of SRA-RiskGate-4B, a 4B model that checks a stablecoin payment before an AI agent sends it and answers approve, hold or reject. For llama.cpp, LM Studio, Ollama and similar.

These files are v2 (2026-11-13), built from the v2 merged weights (full precision: dispute accuracy 90.2% vs 77.8% for v1, unsafe payment approvals 0.35% vs 0.47%). Q8_0 and Q4_K_M were checked on 12 benchmark cases (12 of 12 correct each) but have not yet been run on the full benchmark; the full-benchmark table below was measured on the v1 builds, which stay available at revision v1. The Ollama registry model is still v1 until it is updated.

📄 Paper: preprint on Zenodo · full write-up on Ready Tensor · discuss on alphaXiv

🔁 Verified in the open (November 2026). The companion sra-riskgate library's checks, in front of real agents:

  • Base Sepolia, independently re-checked: an approved 0.01 USDC payment (tx) was read back from the chain and verified, and a 2.00 USDC payment to an unknown payee was held before signing. Flow Executor, an independent agent settlement verifier, re-checked both from the chain (log · Flow Executor's page).
  • Circle Gateway on Arc testnet: one agent task, three proposed payments. One was approved, accepted by Gateway and settled to the payee (agent 0.09 → 0.08 USDC, payee 0.01 → 0.02); an unlimited approval was held and a look-alike payee rejected, neither signed (log).
  • Grok as the paying agent: Grok (grok-4.7) refused both attacks on its own; told to submit everything to the gate, it got APPROVE for the real invoice, HOLD for the unlimited approval and REJECT for the look-alike (logs).
  • Someone else's agent, on mainnet: Minimum Viable Agent's 25.075 USDG loan repayment on Robinhood Chain was pulled by the lender's contract, so there was nothing to sign. It was found on chain by amount and time and checked with verify_settlement: exactly 25.075 USDG to the lender, nothing else moved, allowance left 0 at the block checked. The agent confirmed the result publicly (tx · check · confirmation).

These runs used the library's deterministic checks, not this model.

File Size Notes
SRA-RiskGate-4B-Q8_0.gguf 4.3 GB v2. Closest to full precision; use for disputes
SRA-RiskGate-4B-Q6_K.gguf 3.3 GB v2
SRA-RiskGate-4B-Q5_K_M.gguf 2.9 GB v2
SRA-RiskGate-4B-Q4_K_M.gguf 2.5 GB v2. Payment checks on a laptop

v1 builds on the full benchmark (Q8_0 and Q4_K_M, measured 2026-11-11):

Full precision Q8_0 (4.3 GB) Q4_K_M (2.5 GB)
Payment decision accuracy ↑ 99.5% 99.5% 99.3%
Unsafe approvals ↓ 0.47% 0.47% 0.83%
Legitimate payments blocked ↓ 0% 0% 0%
Dispute outcome accuracy ↑ 77.8% 78.6% 67.4%
Wrongful denials ↓ 26.0% 23.4% 55.3%
Wrongful denials with check_dispute_verdict ↓ 0.85% 0.85% 0.85%

All 2,000 held-out test cases, the exact training prompts, greedy decoding and the benchmark's own score.py; measured 2026-11-11 by the author with Ollama on a T4. Every prediction is published in sra-bench-results. Q8_0 matches full precision. Q4_K_M keeps the payment gate (no legitimate payment blocked; 3 more unsafe approvals than full precision, all where the payer's attestation is missing) but is weaker on disputes: it more often claims a dispute was filed outside the policy's window when it was not, which turns owed refunds into dismissals. That is date arithmetic, so code fixes it: check_dispute_verdict in sra-riskgate brings every build to 0.85% wrongful denials without adding a wrongful refund. Builds other than these two have not been benchmarked.

Which file: Q4_K_M for payment checks on a laptop; Q8_0 (or the full model) for disputes. With any file, run check_dispute_verdict on dispute verdicts.

Run it

Ollama (payment risk-gate system prompt, temperature 0 and an 8K context built in):

ollama run sriram1983007/sra-riskgate

llama.cpp:

llama-server -hf sriram1983007/SRA-RiskGate-4B-GGUF:Q4_K_M -c 8192 --temp 0

LM Studio: search for SRA-RiskGate-4B-GGUF, pick a file, and set temperature to 0.

Use the exact training prompt format from the main model card: the system prompt, then Evaluate this stablecoin payment. with the <context>, <payload> and <tool_results> blocks. Treat any answer that isn't a valid verdict as a hold. In Python, payment_messages() and dispute_messages() in sra-riskgate 1.3.0+ build the prompts for you.

More quantization sizes, including imatrix builds, are available from mradermacher: static and imatrix, and from tinyopsec: SRA-RiskGate-4B-GGUF (Q2_K to F16; builds below Q4 have not been evaluated).

📄 Citation

@misc{ramakrishnan2026srariskgate,
  author    = {Sriram Ramakrishnan},
  title     = {SRA-RiskGate: A Benchmark and a 4B Model for Gating Stablecoin Payments Made by AI Agents, and the Checks the Model Still Needs},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23715017},
  url       = {https://doi.org/10.5281/zenodo.23715017}
}

Full write-up: Ready Tensor.


License

Apache 2.0

Downloads last month
920
GGUF
Model size
4B params
Architecture
qwen3
Hardware compatibility
Log In to add your hardware

4-bit

5-bit

6-bit

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for sriram1983007/SRA-RiskGate-4B-GGUF

Quantized
(5)
this model

Dataset used to train sriram1983007/SRA-RiskGate-4B-GGUF

Collection including sriram1983007/SRA-RiskGate-4B-GGUF