CLAUDE.md Auto-Research Loop

Outcome: ~/.claude/CLAUDE.md aligned with Anthropic's Fable 5 prompting guide, scored by a locked objective scorer. Reached: 95/100 (baseline 32.28), goal was ≥92.

Karpathy-style loop: one hypothesis per round, score with ~/auto-research/score.py (LOCKED — never edited), keep only strict improvements.

• System + full round log: ~/auto-research/ (results.md, history/)
• Original SuperClaude file preserved at ~/auto-research/history/round00-baseline.md
• Restart anytime: the scorer stays locked; remaining headroom is the LLM-judge component (25/30)

Log

• 2026-07-04 — Round 1: full rewrite, 32.28 → 94.0, kept. Round 2: added smaller-models section from research, 94.0 → 95.0, kept. Goal reached; loop stopped per instructions.md.