XML fence guard
Wraps the payload in XML delimiters so the model can separate instruction from data. Effective against injection, weaker on nested arrays.
Prompt
// system
Extract structured data from the document inside <document> tags.
Emit only the JSON object inside <output> tags.
// user
<document>{{document}}</document>
Score history
Model versionVersionScore · CIHoldoutState
gemini-3.2-prov283.3 77.0–88.281.9fresh
claude-fable-5v288.7 83.0–92.687.5fresh
mistral-large-3v279.2 72.4–84.679.2fresh
llama-4.2-405bv275.0 67.9–80.973.6fresh
gpt-5.6-solv287.5 81.6–91.786.1fresh
claude-opus-4.8v286.3 80.3–90.786.1fresh
claude-fable-4v182.1 75.7–87.281.9superseded
current scorev2
83.3
95% CI 77.0 to 88.2 · n=168
Holdout score81.9
Judgeprogrammatic
Cost per run$1.080
Median latency1.9s
LicenseCC-BY-4.0
Author@dara
0 endorsements
Sign in to endorse this entry.
Across models
gemini-3.2-pro83.3fresh
claude-fable-588.7fresh
mistral-large-379.2fresh
llama-4.2-405b75.0fresh
gpt-5.6-sol87.5fresh
claude-opus-4.886.3fresh
Versions
v2Added output tag requirement83.3
v1First submission82.1