LAB / RAG DEBUGGER

Inspect what the model actually saw.

Compare retrieval strategies using synthetic policy documents. Scores are illustrative, not production claims.

01 / QUERY REWRITE

AI agent high-impact tool authorization approval controls

Added domain terms · preserved intent

02 / RETRIEVE

Hybrid + rerank

ACL filter · freshness filter · top 12

03 / CONTEXT

3 evidence chunks

2,146 tokens · 3 sources

Retrieved evidence

#1Tool Authorization Standard0.98

High-impact actions require resource-scoped authorization independent of model intent.

rerank +0.05

#2Human Approval Policy0.96

Destructive or externally consequential actions require explicit human approval.

policy / synthetic

#3Agent Operations Guide0.91

Every tool invocation must carry request identity, policy version, and an auditable result.

policy / synthetic