Deterministic rules
Fast checks for known patterns and policies.
AURA Shield
Auditable prompt screening before the model.
AURA SHIELDPROMPTS SCREENED BEFORE THEY REACH THE MODEL.
01 / THE OVERVIEW
Screen prompts before they reach the model.
AURA Shield is an auditable gateway that screens prompts before they reach an LLM app. Deterministic rules, TF-IDF similarity, and LLM semantic analysis run against a versioned constitution.
Every decision ends in allow, review, or block.
“Rules, similarity, and LLM review, with allow, review, or block.”THE IDEA
02 / THE PROBLEM
Prompt filters are often opaque.
A blocked prompt is hard to trust when nobody can explain why it was blocked or which policy applied.
03 / THE SOLUTION
A gateway with a versioned constitution.
AURA Shield combines deterministic rules, TF-IDF similarity, and LLM semantic analysis, then enforces allow, review, or block. The constitution is versioned so decisions can be traced.
TF-IDF similarity
Retrieval-based comparison against known prompts.
LLM review
Semantic analysis on the cases that need it.
04 / THE IMPACT330 prompts.
330 prompts.
210 tests passing.
The benchmark has separate dev, adaptation, and held-out splits, and 210 automated tests pass.
The cover is a real screenshot of the live AURA Shield console.
NEXT IN THE COLLECTION