Claim-level attribution scoring
Decomposes an answer into atomic claims and scores each one against the retrieved context, rather than scoring the answer as an undifferentiated blob.
An evaluation harness for grounded generation.
Each of these exists because something was broken without it. Nothing here is on the list to fill out a comparison table.
Decomposes an answer into atomic claims and scores each one against the retrieved context, rather than scoring the answer as an undifferentiated blob.
Fixed seeds, pinned judge versions and content-hashed datasets, so a regression in your score means a regression in your system.
A single command, machine-readable output, and a non-zero exit code when attribution drops below your threshold.