AuthorizationPairs

A compact benchmark for deciding whether a request is authorized by its policy and evidence. A common flat schema covers two distinct tracks.

Track	Development	Test	Total	Reference
Language	40	160	200	Expert-reviewed, author-confirmed labels
Native policy	379	1,215	1,594	Official OpenFGA / Cedar engines
Total	419	1,375	1,794	Two reference types

Load

from datasets import load_dataset
cases = load_dataset("YOUR_DATASET_REPO", split="test")
language = cases.filter(lambda row: row["track"] == "language")

Fields

• id: stable case identifier.
• group_id: a language scenario or an exact native policy-and-query group.
• track: language or native.
• domain: language domain or native engine.
• input: JSON string of the original public model-facing input.
• label: ALLOW or BLOCK.
• ambiguous: flags the two strict language-interpretation ambiguities.
• reference_type: separates expert-reviewed language labels from engine labels.
• source: public source identity and pinned revision, or synthetic provenance.
• input_sha256: SHA-256 of the UTF-8 input string.

Send only input to a model. Labels, grouping fields and reference metadata are for evaluation, not model input.

Language track

Fifty scenarios each contain four related contexts for one fixed operation: two authorized and two unauthorized. Ten scenarios are development and forty are test. The material is fictional and was drafted with AI assistance. The authors report independent review by three security/access-control annotators and confirm that the final labels agree case by case with the frozen scorer. Individual annotation votes are not included in this minimal release.

Two contexts retain the published ALLOW labels but remain unresolved under a stricter encoded reading of the approval issuer. They are marked ambiguous=true; reporting results with and without these cases is supported. These cases should not be silently relabeled. Scenarios and their four contexts are correlated.

Native-policy track

There are 1,548 distinct OpenFGA requests and 46 Cedar requests. References were checked with official native engines against public source assertions. Source applications and shared exact policies define the split. The dataset includes 23 groups with both reference decisions for the same policy and query, of which 22 are in test. Most native cases are individual permission queries, not pairs. Native context may vary in several fields. No English paraphrase replaces the native policy, and no business operation is executed by a classification script.

Sources:

• OpenFGA sample stores
• Cedar examples

Evaluation

Report valid-decision accuracy, authorized approval rate, unauthorized approval rate and unavailable/invalid output rate together. An invalid response does not receive accuracy credit merely because it prevents approval. Contrast-group success requires all decisions correct in a complete group containing both labels.

A minimal API baseline measures decisions only. It does not reproduce native application effects or every reviewer-specific interface. Public-source policies may overlap model training data. Case count is not a count of independent workflows.

License and provenance

The release uses Apache-2.0. Native examples retain their upstream license and attribution in NOTICE and licenses/. manifest.json records case counts and checksums. No credentials, private project paths, model predictions or author identities are embedded in the exported data. The hosting account name remains visible as part of the platform namespace.