When workflow completion is not enough: a synthetic study of automated action integrity
Description
A deterministic synthetic study of application-level automated-write integrity using official Temporal and LangGraph components and Promptfoo assertions. Six scenarios per engine, with guarded and unguarded applications, produced 24 evaluations: 16 passes, 8 expected unguarded violations and zero evaluation errors. Repeatability was checked with fresh state in the same installed dependency environment.
The study examines idempotency, concurrency, permission revocation and audit consistency. Findings concern the tested application configurations; they are not vendor vulnerabilities, safety certification or a benchmark ranking.
This work was self-funded. No external funding was received. AI assistance was used in implementation and writing. No independent peer review is claimed.
Article/text licence: CC BY 4.0. Selected original synthetic harness: MIT; third-party components retain their own licences.