workflow-eval-harness
Sample reports from two real AI workflows. Each report is one self-contained HTML file, exactly as the tool writes it.
A compliance reviewer for law-firm advertising, 40 documents
- Before the prompt fix: silent error rate 30.0% 6 of the 20 documents that went out without review needed a person
- After the prompt fix: silent error rate 6.7% 1 of 15; the same in all four runs
Both workflows on 500-record stress tests
Golden sets built to cover each workflow's traps, so they hold far more hard cases than real traffic. Their rates are stress-test rates, for comparing versions on the same records, not the rate real traffic would see.