Extract then compress
Pulls key claims verbatim first, then compresses only those. Cannot introduce material that was not extracted.
Prompt
// system
Step 1: list the key claims verbatim. Step 2: compress only those claims into a summary.
// user
{{source}}
Score history
Model versionVersionScore · CIHoldoutState
gpt-5.6-solv186.6 79.1–91.785.4fresh
claude-fable-5v187.5 80.1–92.487.5fresh
mistral-large-3v178.6 70.1–85.277.1fresh
claude-opus-4.8v184.8 77.0–90.383.3fresh
llama-4.2-405bv173.2 64.3–80.572.9fresh
gemini-3.2-prov182.1 74.0–88.181.3fresh
current scorev1
86.6
95% CI 79.1 to 91.7 · n=112
Holdout score85.4
Judgemodel
Cost per run$0.792
Median latency2.5s
LicenseCC-BY-4.0
Author@ilya
0 endorsements
Sign in to endorse this entry.
Across models
gpt-5.6-sol86.6fresh
claude-fable-587.5fresh
mistral-large-378.6fresh
claude-opus-4.884.8fresh
llama-4.2-405b73.2fresh
gemini-3.2-pro82.1fresh
Versions
v1First submission86.6