| --- |
| language: |
| - en |
| license: apache-2.0 |
| tags: |
| - ai-agents |
| - evaluation |
| - benchmarking |
| - agent-verification |
| - safety |
| --- |
| |
| # AffixIO Agent Eval Benchmarks (Dataset repo) |
|
|
| Offline safety benchmark: correct tool selection (12), prompt injection |
| resistance (12), permission boundary handling (11), email classification |
| accuracy (12), safe action decisions yes/no/review (13), **agentic |
| payments (8)** — 68 cases total. |
|
|
| ```bash |
| python run_bench.py |
| ``` |
|
|
| Reference result (v1.1.0): overall 0.941 — tool_selection 1.000, |
| injection 1.000, boundary 0.818, classify 1.000, action 0.923, |
| pay 0.875. |
| Release gate: overall >= 0.90 and no suite < 0.80. |
| |
| Files: `cases.py` (56 cases), `run_bench.py` (runner + gate). |
| The runner imports `../../03-agent-tools/tools.py` when present, else an |
| inline copy of the same policy. |
|
|
| ## Version |
|
|
| - v1.0.0 — initial 56-case benchmark. See README.md. |
|
|
| ## Links |
|
|
| - Product: https://www.affix-io.com |
| - SDK: https://www.npmjs.com/package/affixio |
| - API docs: https://www.affix-io.com/docs/ |
| - Agentic Pay Kit: https://www.affix-io.com/agent-trust/ |
|
|
