BuildMidData / MLAuto-graded

Flag claims no source actually supports

Making an automated fact-check tolerate real wording, and point at the right source, is what gets measured.A retrieval-augmented system answers by reading retrieved source snippets and writing sentences, and the whole promise is that every sentence is grounded in those sources. In practice the model drifts — it paraphrases something the sources never said, blends two facts, or invents a plausible detail. Build the guardrail that reads each claim back against the sources and flags the ones that are not actually supported.

35 min

python

Start lab

You're done when

The checker runs over claims and sources without crashing.
It returns a result in the documented shape, with a grounded or ungrounded verdict per claim.
Matching is robust to normalisation differences — casing, whitespace and punctuation do not by themselves make a supported claim look unsupported.
A claim with no sources at all is reported as ungrounded rather than passing by default.
Claims genuinely supported by the sources are not flagged.

What it measures

AI evaluation and QABuilding an automated check for groundedness, the property RAG systems are supposed to guarantee.

RAG and retrieval systemsUnderstanding how models drift from their sources and which drift a checker can realistically catch.

Defaulting to safeTreating absent evidence as ungrounded, so the failure mode is a flagged claim rather than a passed hallucination.

Your environment

Terminal + editorA full shell in the browser

Claude CodePre-authenticated, no key setup

Isolated microVMYours alone, wiped afterwards

Start this lab

35 min

Start labYou'll sign in to Anthropos to start. The lab opens in a private cloud workspace with Claude Code ready to use.

AI Labs is part of Anthropos Workforce. Not on your plan yet? Book a demo