Comprehensive prompt refinement: verbatim numbers, template-lock fix, tighter targets

[neoneye]

Summary

Bundles all proven improvements from analyses 69-73 into one definitive PR:

1. Verbatim-numbers on consequences AND options: "Use numbers only when the project context provides them directly — do not calculate, derive, or estimate figures." Closes the arithmetic-derivation loophole.
2. Template-lock fix (top issue from analysis 40): Replace "the three options leave unaddressed" with "the proposed options collectively do not resolve" in review_lever. Applied to field description, section 4, AND section 6.
3. Positive framing: Replace "Do NOT include 'Controls ... vs.'"
4. Consistent targets: 2-3 sentences, one-sentence options/review, aligned across all field descriptions and system prompt sections.
5. Section 5: "NO calculated, derived, or estimated figures — use only numbers that appear verbatim in the project context"

Supersedes #471, #473, #475, #477, #478.

Test plan

• Verify gpt-oss-20b completes 5/5 plans
• Verify haiku review template lock drops from 70-94%
• Verify fabricated numbers ≤ baseline (23)
• Verify real plan numbers still cited when appropriate
• Assessment by GPT5.4 confirms convincingly better than analysis 40

🤖 Generated with Claude Code

[neoneye]

Self-improve iteration — analysis 74

Verdict: CONDITIONAL

Win:

• gpt-4o-mini template lock: 5/16 → 0/17 (genuine improvement)
• 100% call success rate (up from 97.1%)

Regressions:

• qwen3-30b new template lock: 0/15 → 7/20 ("collectively do not resolve" is itself a copyable phrase)
• gpt-oss-20b fabricated numbers: 0 → 5 (Section 5 abstract prohibition activated number-generation)
• gpt-oss-20b template lock shifted and increased: ~10/17 → ~15/17

The fundamental problem: replacing one copyable phrase with another. Any sentence-structural phrase in the field description becomes a template for small models. The assessment cites OPTIMIZE_INSTRUCTIONS lines 86-92 — the PR violated its own documented anti-pattern.

[neoneye]

Closing — replaced one template lock with another. See iteration_insights.md in analysis 74.