VAIG (VALO AI Integrity Gateway) — ACS-compatible runtime governance layer, Apache 2.0

[nsolland]

Hi,

We're building VAIG — an open-source runtime governance layer for LLM
inference, designed for EU AI Act Annex III compliance and
insurance-grade auditability. Apache 2.0.

How VAIG maps to ACS

ACS principle	VAIG
Runtime Interception	Probabilistic distrust scoring at every inference
Proof of Control	SHA-256 hash-chained WORM audit log, append-only
Mandatory Halt	Binary HALT state, formally verified (TLA+, 4.78M
	states, 0 counterexamples)
Contextual Guardrails	DistrustEngine L0–L4 (TRUSTED → MONITOR → WARN →
	DEGRADE → HALT)

Key design principle

VAIG sits architecturally outside the model it monitors — it
cannot be influenced, overridden, or manipulated by the system it
validates. This is the same principle that makes Safety Instrumented
Systems (IEC 61511) reliable in industrial process control.

Status

• Formally verified halt mechanism (TLA+ model-checked)
• WORM audit log producing legally defensible audit trails
• Validated on MedQA and LEDGAR benchmarks
• Designed for deployment on sovereign European infrastructure

We'd like to be listed as a compatible implementation alongside
LangChain/CrewAI. Happy to contribute documentation or a reference
integration guide.

— Njål Solland, VALO Research Group
https://github.com/nsolland/VAIG
pip install vaig

[Keesan12]

This is a strong direction. Runtime governance only earns trust when it is present at the moment a real decision gets made, not as a report someone reads after the damage is done.

I'm building MartinLoop from the coding-agent side of the same problem. We have learned that a team does not just need a policy result; it needs a plain answer after each run: what did the agent try, what was allowed, what changed, did it pass verification, and can it be undone. That is how AI writing code becomes AI doing trusted software work.

Would be glad to compare notes on the smallest common contract a governance layer should expose. We are building in the open at https://github.com/Keesan12/martin-loop. Reach me at support@martinloop.com if a direct exchange is easier.