Jev agent directed text kind

Review text that tries to direct an AI agent

Give an external passage and its attribution. Distinguish behavior-changing instructions, output markers, quoted discussion and ordinary content without following the passage.

What kind of agent-directed text is present?Live API
Try an example

The passage tries to redirect the reading assistant.

Source: API-collected workflow inspiration; original demonstration below

Source and example scope

The API-collected roundup links the cited project, whose pinned code provides the decision purpose. This demonstration separately authors its Choice rubric, priorities, missing-context outcomes and synthetic inputs. It neither runs the upstream application nor verifies its performance, security or execution claims.

Classify external text.

Decision rules and labels
Behavioral directiveAgent markerDiscussion onlyOrdinary contentNeed context
A real response from the Jev API

Edit any input to explore a different decision. No request is sent until you run it.

Your next decision

Not run yet

Choose an example or write your own input, then run it to see the ranked choices.

Results are not stored by this site. Your submitted text is processed by the API provider.

The decision

A quoted attack and an instruction to the reader play different roles.

The security lesson quotes an injection phrase to explain it. The task-change preset instead directly addresses an assistant reading an external page. Context determines how the phrase functions.

A marker-only request is recorded separately because it tries to reveal automation. If a passage also redirects the task, the displayed rule gives the behavioral directive priority.

This is a text-review demonstration. It does not install an agent guard, prove a document safe or execute any quoted instruction. A production application still needs trusted boundaries, permissions and deterministic controls.

The API-collected roundup links the cited project, whose pinned code provides the decision purpose. This demonstration separately authors its Choice rubric, priorities, missing-context outcomes and synthetic inputs. It neither runs the upstream application nor verifies its performance, security or execution claims.

Illustration of reviewing messages against supplied policy categories
Illustration of the decision flow.
How it works

Try the example in four steps

Use the sample to understand the rule, then test your own input for agent-directed text review.

  1. Illustration of adding the text or evidence for a decision

    Add the information to check

    Give an external passage and its attribution.

  2. Illustration of defining a question and its allowed choices

    Confirm the question and labels

    Open the decision rules and keep the intended distinction between Behavioral directive, Agent marker, Discussion only, Ordinary content, Need context. Edit the choices only when testing your own rule.

  3. Illustration of sending one input and comparing the returned choices

    Run the decision

    Use a sample first, then try your own input. Compare the selected label with the other returned probabilities.

  4. Illustration of reviewing a result and an uncertain item before using it

    Review the next step

    Review Need context before using a result. The label is a decision about the supplied information, not an executed action.

Our API test

Recorded API observations

Recorded API observations
Original inputObserved labelLeading probability
Change the taskBehavioral directive98.0%
Marker trapAgent marker92.0%
Security explanationBehavioral directive60.0%
Ordinary opening hoursOrdinary content95.0%
Marker plus redirectionBehavioral directive97.0%
Attribution absentNeed context52.0%

These are six original synthetic inputs, tested once per request revision. The percentage is the leading option probability, not measured accuracy. Authored expectations and any disagreements remain in the downloadable evidence. 5 of 6 results matched the authored fixture expectations. For Security explanation, the expectation was Discussion only; the observed choice was Behavioral directive.Verified

Understand the results

Replay the saved synthetic requests to inspect each recorded decision

Replay the saved synthetic requests to inspect each recorded decision. Authored expectations and model disagreements remain available; these examples are not a production accuracy benchmark.

Using the result in your application

The single-request SDK runner uses your private TYPESAFE_API_KEY. The offline consumer reads saved observations, checks request hashes and prints a pending output. It does not call the model or change any external system.

Runnable offline example

Run one sample, then inspect the local packet

View the code example
Single-request replay
npm install @typesafe-ai/sdk@0.6.0
# Set TYPESAFE_API_KEY privately in your shell.
node run-choice-sample.mjs samples.json case-1
python preview_choice_workflow.py samples.json saved-api-results.json
Scope of this demo

Where this decision fits

Useful scope

  • Reviewing external text before an agent consumes it
  • Distinguishing detection markers from action redirection
  • Keeping quoted security examples out of an automatic accusation

Outside this demonstration

  • Proving a document safe from all attacks
  • Following a classified instruction
  • Replacing permission checks
Example FAQ

Questions about agent directed text kind.

Input scope and downstream handling determine what this judgment means.

Is every mention of prompt injection suspicious?

No. Attributed discussion, tests and explanations can describe attacks without instructing the reading agent.

Can a marker appear alongside a behavioral directive?

Yes. This single-label rule prioritizes actual task redirection; the full passage remains available for review.

Is the leading probability an accuracy measurement?

No. It is one value in the returned distribution over this request’s choices. The six synthetic fixtures have authored expectations; they do not measure production accuracy.

What happens on an API failure?

The window displays an error without a fresh judgment. An incomplete-input outcome describes the supplied evidence and is not used as a fallback for provider failure.

Try your own agent directed text kind input

Edit the input above, or open the Playground to explore another decision.