| --- |
| license: apache-2.0 |
| pretty_name: governed-agent-bench |
| task_categories: |
| - other |
| tags: |
| - governance |
| - agents |
| - evaluation |
| - receipts |
| - rollback |
| --- |
| |
| # governed-agent-bench v0 |
|
|
| This dataset is the immutable public mirror of |
| [`szl-holdings/a11oy@1b40fcbe0f65c1ad1e07776f83aa01abb067e864`](https://github.com/szl-holdings/a11oy/tree/1b40fcbe0f65c1ad1e07776f83aa01abb067e864/benchmarks/governed-agent-bench). |
|
|
| It measures five governability axes: |
|
|
| 1. fail-closed behavior; |
| 2. non-increasing authority across delegation; |
| 3. false-success rejection; |
| 4. receipt completeness; and |
| 5. rollback discipline. |
|
|
| ## Evidence labels |
|
|
| - Corpus: **SAMPLE** |
| - Scores: **COMPUTED** |
| - Receipt verification: **STRUCTURE_ONLY** |
| - Cryptographic verification: **false** |
| |
| The reference result proves that the evaluator and known-good fixture close |
| their deterministic contract. It is not a model-quality or production claim. |
| The public leaderboard contains zero eligible model submissions until an exact |
| submission is evaluated and published with its receipt. |
| |
| ## Reproduce |
| |
| ```bash |
| python score.py submissions/reference-conformance.jsonl --strict |
| ``` |
| |
| The canonical source, schema, evaluator, reference submission, result, and |
| publication manifest are all included in this dataset revision. The companion |
| Space is <https://huggingface.co/spaces/SZLHOLDINGS/governed-agent-bench>. |
| |
