Official datasetProvider-published evidence
WorkflowEvals Invoice Processing
A 150-case dataset for invoice approval and fraud-triage decisions under three policy settings.
The official EvalSafe dataset separates workflow cases, question instances, and model run results so evaluators can inspect both source inputs and action-level outputs.
- Publisher
- TypeSafe AI
- Published or submitted
- Not specified by the source
- Last verified
- October 8, 2026
Type Safe catalogs sources and evidence boundaries. Provider datasets and third-party preprints do not become an independent leaderboard.
Key facts
- Cases
- 150
- Question instances
- 6,874
- Published run results
- 1,350
- License
- Apache-2.0
Task and scope
Invoice approval and fraud triage under three policies.
Dataset schema
cases · questions · run_results · 8,374 rows
Evaluation metrics
exact_actions · primary_action
Dataset license
Apache-2.0
Method and usage
- The workflow asks models to approve, reject, flag, or escalate invoice cases under three policy configurations.
- Published results can be compared by exact action-set agreement and primary-action agreement where applicable.
Limitations and evidence boundaries
- Policy-conditioned synthetic or curated cases do not cover every accounting system, fraud pattern, or jurisdiction.
- Viewer row totals combine different schema groups and must not be interpreted as independent cases.