A grounding enforcement layer for AI agents tested against NeMo, Guardrails AI, and Patronus Lynx

I’m building AletheionAGI, a grounding enforcement layer for AI agents and RAG systems.

The goal is simple: let AI use memory and retrieval, but do not allow unsupported or unauthorized claims to pass through.

Aletheion sits between retrieval/evidence and the final answer and checks whether the available evidence actually supports the claims being produced. If support is insufficient, it can fail closed.

I recently ran the same frozen black-box protocol against:

• AletheionAGI

• Guardrails AI

• NVIDIA NeMo Guardrails

• Patronus Lynx

Final observed results:

AletheionAGI — 9 PASS · 0 FAIL · 1 SKIPPED · 0 ERROR
Guardrails AI — 4 PASS · 2 FAIL · 4 SKIPPED · 0 ERROR
NVIDIA NeMo Guardrails — 5 PASS · 2 FAIL · 3 SKIPPED · 0 ERROR
Patronus Lynx — 5 PASS · 1 FAIL · 4 SKIPPED · 0 ERROR

The protocol covers cases such as unsupported claims, poisoned or revoked evidence, authorization leakage, namespace isolation, and fail-closed behavior.

The single AletheionAGI SKIPPED case required a requester/label denial policy that was not present in the test setup, so the protocol correctly refused to manufacture a PASS or FAIL.

This is a black-box proof of concept, not a claim of universal superiority.

Protocol and implementation are available for inspection and testing.

https://aletheionagi.com

I’m especially interested in feedback from people working on RAG, AI agents, grounding, guardrails, and LLM security.