API guardrails

A guardrail is a call to one of your judges from your own code. You send one piece of content, the judge grades it against its rubric, and the call blocks until the verdict comes back, so your system can act on it at once: hold the content back, hand it to a person, or record the score. Use a judge you have calibrated first, so its verdicts match your reviewers.

Making a call

from dataframer import Dataframer

client = Dataframer()  # uses DATAFRAMER_API_KEY
evaluation = client.dataframer.judges.evaluations.create(
    judge_id=JUDGE_ID,
    input="Summarize the attached policy in two sentences.",
    output="The policy covers water damage from burst pipes but excludes flooding.",
)

Send the content as a plain input/output pair, or as a chat transcript in messages with the candidate output.

Reading the verdict

The verdict carries one result per rubric dimension, scored from 0 to 1. It also carries the judge’s confidence, its reasoning, and an aggregated overall score. Every call is stored. In the app, open Evaluations → Guardrail API calls to read them back, or list them through the API.

Next steps

API reference

Full payload, file attachments, and error codes

Get started

Capture & Correlate

User Journeys & AI

Findings

Human review operations

Judges and Evaluations

Data Generation

Programmatic Access

Making a call

Reading the verdict

Next steps

API reference

​Making a call

​Reading the verdict

​Next steps

API reference

Making a call

Reading the verdict

Next steps