Skill+evals: number guard and round 7

[davehallmon]

What this changes

• SKILL.md: adds a guard to "Use the user's details": derive a number only from facts the user gave, or name the assumption next to the figure. Never present a derived figure as known.
• Round 7 logged in evals/runs/round-7_draft/, including calc-audit/, a blind audit of every derived number in rounds 6 and 7.
• evals/README.md and CHANGELOG.md updated.

Round 7 results (same 11 problems, same baselines)

• Calculation audit: unsupported derived numbers fell from 9 to 3. No arithmetic errors in either round. Two of the three left are effort estimates like "it costs an hour."
• Honesty: 4.73, the highest of any round.
• Overall: 3–1–7, gap 0.0. The Skill's mean (29.5) is unchanged from round 6; the baseline files scored higher this time. Judge noise this round was 1.64 points per pair, the highest yet. Five of the seven losses were by one point.
• Fit: 4.18 (round 6: 4.09). Plain Claude: 4.73.
• Rule checks: 10 of 11 pass. The receptionist Summary ran to four sentences.

Checklist

• I read CONTRIBUTING.md.
• I used active voice and short words.
• I kept SKILL.md under 500 lines (127).
• I updated README.md if needed (left for the 1.2.0 release PR).
• I updated CHANGELOG.md.
• I tested the Skill on real problems (eval round 7: 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