[JacobStephens2]

Summary

Research for #429 (multi-family review) and #428 (more Harnesses). Docs only.

 docs/research/
+├── multi-family-review.md        # another family reviewing a Run's work, and mixing families in coding; searched through 5 Oct 2026
+├── model-strength-for-review.md  # whether AA's Intelligence Index, or anything, says which model is the stronger reviewer
+└── harness-candidates.md         # Muse Code, Grok Build, MiMo Code and agy, run headless on one machine

Evidence

• Before: add support for more harness/model pairs #428 and Multi-family review: have another model family review a Run's work #429 rested on recollection and a pasted summary, which read Greptile's chart as a ~9-point cross-family gain.
  After: every claim cites its source, and the numbers the decisions lean on were checked against the originals:
  • The July paper (arXiv:2607.21656): its abstract and its artifact's statistics file. Cross- vs same-family review of GPT drafts was 10 tasks vs 4, p_BH = .245.
  • Greptile's chart markup. Coral marks Opus and a dashed bar marks the same model, so the cross-family gains are +2.0 and +3.2 points, not ~9.
  • SWE-Review Table 2.
  • AA's methodology page (v4.3.2, ten evaluations) and leaderboard data. Opus 5.5 at medium scores 51.2 and GPT-6.1 Sol at xhigh scores 51.0.
  • Anthropic's Opus 4.7 page, which gives a January 2026 training cutoff.
  • The abstracts of arXiv:2610.01471, 2610.01023, 2609.10969, 2609.22512, 2609.20812 and 2609.34198.
• The Harness facts come from headless runs. Each note marks how every fact was confirmed: by test, help text, docs or source.

Merge Danger

Door: two-way

Blast Radius: none

Docs only: no code, prompts or skills change.

🤖 Generated with Claude Code