fix: remove 'core tension' template lock from field description

[neoneye]

Summary

The phrase "name the core tension" in the review_lever field description (line 111) and system prompt section 4 (line 228) caused ~100% "The tension is between X and Y" opener lock across all plans for llama3.1 and haiku (confirmed in iter 39, analysis 39).

• Field description: "name the core tension, then identify a weakness" → "identify the primary trade-off this lever introduces, then state the specific gap"
• Section 4 header: same rewrite
• Section 6: "under 2 sentences (aim for 20-40 words)" → "one sentence (20-40 words)" — removes ambiguity
• OPTIMIZE_INSTRUCTIONS: document field-description template lock as a known failure mode

The three section-4 examples already use domain-specific openers (not "The tension is..."), so removing the structural cue from the instruction text aligns instruction with examples.

Test plan

• Run self_improve iteration (7 models × 5 plans)
• Verify "The tension is between X and Y" opener drops from ~100% to <40% for haiku
• Verify llama3.1 reviews use diverse openers instead of uniform "tension" pattern
• Verify success rate stays at 100%
• Verify no regression in silo/hong_kong template lock (should remain at 0%)

🤖 Generated with Claude Code

[neoneye]

Self-Improve Iteration 40 — Verdict: YES ✅

"Tension" template lock eliminated.

Metric	Before (iter 39)	After (iter 40)
"Tension" lock (haiku hong_kong)	75%	0%
"Tension" lock (llama3.1)	~100%	~0%
Haiku call success rate	73.3%	86.7% (+13.4pp)
Overall call success rate	95.2%	97.1% (+1.9pp)

Secondary lock emerged

• "None/All three options" at ~85% for haiku — caused by "the three options leave unaddressed" in new field description
• Same template-lock migration mechanism, milder form — target for next iteration

Minor cleanup needed

• LeverCleaned.review field description (line 212) still says "core tension" — not sent to LLM, cosmetic only

Data: PlanExe-prompt-lab/analysis/40_identify_potential_levers/