jev screen

Judge text before an AI agent reads it: is it trying to hijack the agent, does it have substance, is it relevant to the task.

jev screen [text] [--purpose <text>] [options]

When to use

On every untrusted input an agent consumes: fetched web pages, emails, tickets, uploaded documents, third-party tool output. It costs less than reading the page.

Options

Option	Meaning	Default
[text]	Text, @file, or -. Reads stdin when omitted and piped.	stdin
-p, --purpose <text>	What the consumer is trying to do. Enables the relevance check and the skip action.
--block-at <p>	Injection probability that triggers block	0.75
--review-at <p>	Injection probability that triggers review	0.25
--fail-on <list>	Exit 2 when the recommendation is block, review, or skip. none to always exit 0.	block

Output

Field	Meaning
probabilities.injection	The text contains instructions aimed at an AI agent
probabilities.substance	The text has real content, not an error page or boilerplate
probabilities.relevance	The text is useful for --purpose (null without a purpose)
recommendation.action	pass, review, block, or skip
recommendation.reason	Which threshold fired

block and review come from the injection probability; skip from low substance or relevance. The recommendation is advisory: jev never blocks anything itself.

Example

curl -s https://shop.example/sale | jev screen --purpose "summarize the products"

BLOCK  injection probability 0.99 >= block threshold 0.75
injection 0.99 · substance 0.97 · relevance 0.97

# Only feed the page to the agent if it is safe and relevant
if printf '%s' "$page" | jev screen --purpose "$task" --fail-on block,review,skip -q; then
  printf '%s' "$page" | my-agent --context -
fi

Pattern: guardrails cookbook.