Pairing weakest anchor agent with best evolver against strongest anchor with worst evolver, the strong agent still leads by 18.6 to 35.2 pp on every benchmark

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Qwen3-235B leads as evolver on SWE-bench with 8.2 pp harness-updating gain but ranks last on MCP with 0.6 ppfinding0.773
  Illustrates benchmark-dependent reshuffling of evolver rankings, no evolver dominates across all substrates
• Within-agent spread across seven evolvers is at most 5.1 pp (Qwen3-235B on MCP), small against the 36.0 pp gap between Opus and Qwen3-235B base capabilitiesfinding0.771
  Demonstrates that post-evolution score is dominated by agent base capability, not evolver identity
• Capability budget should be allocated to the task-solving agent rather than the evolver, since harness-updating varies by at most 3.1 pp across evolversclaim0.761
  Primary design recommendation derived from harness-updating flatness finding
• Qwen3.5-9B evolver achieves highest harness-updating gain on SkillsBench (3.8 pp), exceeding Claude Opus 4.6 (2.3 pp) and Qwen3-235B (1.5 pp)finding0.746
  Case demonstrating that model scale does not predict harness-updating quality
• Dense but off-task anchors yield high ρd AND high dr; behavior does not improve, consistent with mismatch dominating Sfinding0.744
  E3 negative control validating that both ρd AND dr must be favorable for S to exceed Sc
• Math and code tasks show strongest mid-layer anchoring on LLaMA (S ≈ −1.65 at layers 8-12)finding0.744
  Task-specific E3 finding showing compositional reasoning requires deeper processing
• Peak anchoring Sbmax and normalized area AUSN correlate with per-item success and internal shot midpoints θ50, providing a geometry-to-behavior bridge.claim0.740
  Main interpretation of E3.
• Feature pair A/1/3949 and B/1/3321 have activation correlation 0.98 but negative logit weight correlation, firing on PLOSOne journal abbreviationsfinding0.737
  Demonstrates that activation similarity can diverge from logit weight similarity due to interference