Open-Source Tool 'Evidence Graph' Released to Enforce 100% Skill Guideline Compliance in AI Code

Wrtn Labs has released the Evidence Graph tool, which enforces full compliance of AI-generated code with Skill guidelines. This addresses the 'The Compliance Gap' problem, where AI understands specifications but ignores certain constraints while completing tasks.

Core Mechanism

Compliance status or reasons for non-application of each directive are explicitly marked with tags (@evidence, @evidenceExclude) within the code to generate a reviewable graph. The Checker verifies the existence of answers for required connections, while the Reviewer determines the truthfulness of those answers. Requirements and implementation targets are linked via evidence.config.ts, and the check fails with exit code 1 when running npx evidence upon violation.

Benchmarks and Effectiveness

Benchmark results from building 4 apps with the same engine/model showed that the existing method (Plain) achieved only 51.6~85.5% coverage, with 90% of tokens consumed by repetitive reviews. In contrast, the Evidence Graph method achieved 100% coverage and could be completed with a single review using the tag list.

Support Scope and Features

• Language Support: Supports over 20 languages including TypeScript, Python, Go, Rust, as well as Markdown, Swagger, and Prisma
• Tech Stack: Uses Tree-sitter and a custom parser (no compiler execution)
• License: MIT License (GitHub: wrtnlabs/evidence)