AIGIS: An Open-Source Compiler for AI Governance

Context

Most AI governance tooling assumes a human will read the output, but AI coding agents consume context differently: limited windows, structure-parsing, explicit branching signals. AIGIS v2 was rebuilt around that gap. One command, aigis build "customer support chatbot with order history", compiles a natural-language project description into a deterministic governance brief mapped against NIST AI RMF, OWASP Top 10 for LLMs, and ISO/IEC 42001.

Problem

Developers shipping AI features rarely know which governance obligations apply to them, and existing frameworks are hundreds of pages written for compliance teams. How do you make governance consumable by the AI coding agents that increasingly write the code, in a way that is deterministic, auditable, and free of safety-washing?

Approach

Rebuilt v1 (a content library) into v2 (a compiler): a deterministic layer (39-trigger curated resolver, byte-identical brief assembly verified by SHA-256), a content layer (governance procedures as markdown "fat skills"), and an agent layer (runtime reasoning delegated to the user's coding agent). Development was benchmark-driven: ~170 agent invocations across 8 steps, measuring precision, recall, and F1 for governance-trait classification, with honest disclosure of a recall regression.

Frameworks

Agent-Computer Interface design (SWE-agent, NeurIPS 2024)Thin Harness / Fat Skills architectureNIST AI RMF · OWASP LLM Top 10 · ISO/IEC 42001Benchmark-Driven Development (precision / recall / F1)

Implementation

  • •Deterministic 39-trigger resolver with JSON-validated, schema-checked trigger-to-trait mappings
  • •Overclaim recognizers: regex checks that catch safety-washing patterns and prevent false governance claims
  • •Byte-identical brief generation: the same input always produces the same SHA-256-verifiable output
  • •Dual-registry publishing (npm @aigis-ai/cli + PyPI aigis-cli) with CI-enforced content parity between packages
  • •IDE integration via rules files for Cursor, Claude Code, Copilot, and Windsurf

Outcomes

  • ✓6,000+ total downloads across npm and PyPI with zero paid promotion
  • ✓446 downloads in the first 24 hours of the v2 launch
  • ✓Endorsed by a former Washington Post CTO
  • ✓Governance benchmark: precision 0.847 (+15% over no-tool baseline), F1 0.837 (+6%), rubric score 48.7/50 vs 27/50 baseline
  • ✓Published full benchmark methodology including the recall regression, choosing composite-metric integrity over single-metric wins

Learnings

  • →Content designed for agent consumption first is a different discipline from documentation for humans
  • →Deterministic operations belong in code, domain judgment in content, runtime reasoning in the agent
  • →Honest negative findings build more trust than inflated benchmarks
  • →Launch-day distribution is a product surface: multi-registry parity and release engineering matter