Tier 1 — floor
5/5
184 tokens · 12.753s (representative seed)
Tier 2 — restraint (refusal)
5/5
323 tokens · 3.888s (representative seed)
Tier 3 — distractor resistance
5/5
290 tokens · 3.519s (representative seed)
Tier 4 — multi-passage synthesis
1/5
4096 tokens · 48.798s (representative seed)
Tier 5 — precise extraction
5/5
223 tokens · 2.725s (representative seed)
Tier 6 — long-context needle
5/5
199 tokens · 2.646s (representative seed)