Evals: round 6 (use-the-details rule) - logged, rule reverted

[davehallmon]

What this changes

Logs eval round 6, a test of a SKILL.md rule aimed at Fit, and then reverts the rule because it didn't work. Net effect on the Skill: none. SKILL.md matches main.

Commits:

1. skill: adds "Use the user's details": let a stated number or limit change a check, option, or order, and keep moves doable with the user's resources.
2. evals: logs round 6 in evals/runs/round-6_draft/ and updates evals/README.md and CHANGELOG.md.
3. Revert removes the rule.

Round 6 results (same 11 problems, same baselines as round 4)

• Fit did not move: 4.09 in round 4 and round 6, up on 2 problems, same on 7, down on 2. Plain Claude: 4.73.
• Overall: 6–1–4, gap +0.8 (round 4: +1.6). This is within judge noise; the same baseline files moved 1.45 points on average.
• The rule worked when it worked: turnover worked out that 28% is about 11 exits a year, too few to trust exit-interview themes, and the judge called that the deciding insight.
• It also produced a confident wrong number: onboarding estimated an A/B test at about 4 weeks. The judge put it at 6–7, and noted the run assumed all 200 signups reach step 3.
• Judges' Fit complaints are mostly about coverage, such as the career run not using "8 years of experience." That isn't what a number rule fixes.
• Rule checks passed 11 of 11. OODA was chosen for the first time, so all 11 eligible lenses have now been used.

If you want to keep the rule anyway

Drop the revert commit before merging.

Checklist

• I read CONTRIBUTING.md.
• I used active voice and short words.
• I kept SKILL.md under 500 lines (unchanged after the revert).
• I updated README.md if needed (not needed: no Skill change).
• I updated CHANGELOG.md.
• I tested the Skill on real problems (eval round 6: 11 problems).
• I ran python3 evals/check_rules.py problem-lens/examples/*.md (3 of 3 pass).

🤖 Generated with Claude Code

https://claude.ai/code/session_01X9xscdGXL5xQmmxnwEf9T7

---

Generated by Claude Code