PR #479 + minimal review_lever to avoid template lock

[neoneye]

Summary

Same as PR #479 but with one tweak to fix its biggest regression:

The tweak: Strip review_lever field description to "Critical review of this lever (one sentence, 20–40 words)." — no structural phrase. Section 4 preamble becomes just "A one-sentence critical review. Examples:". The three existing examples teach the style without a copyable template.

Why: PR #479's "the proposed options collectively do not resolve" caused qwen3-30b to lock (0→7/20) and worsened gpt-oss-20b (10/17→15/17). Every structural phrase becomes a fill-in-the-blank template for small models.

From #479 (unchanged): verbatim numbers on consequences+options, positive framing, consistent 2-3 sentence targets, section 5 prohibition.

Test plan

• Verify haiku review template lock drops from ~85% baseline
• Verify qwen3-30b does NOT gain a new lock (was 0→7/20 in Comprehensive prompt refinement: verbatim numbers, template-lock fix, tighter targets #479)
• Verify gpt-oss-20b lock doesn't worsen
• Verify no content regressions
• Assessment by GPT5.4 confirms better than analysis 40

🤖 Generated with Claude Code

[neoneye]

Self-improve iteration — analysis 76

Verdict: CONDITIONAL

Template lock BROKEN — all three models fixed:

• qwen3-30b: 100% locked → 0% (confirmed across all 101 levers)
• haiku: ~85% "All three options" lock → ~0% (diverse, domain-specific reviews)
• gpt-oss-20b: pattern present → absent

The key insight: Minimal field description "Critical review of this lever (one sentence, 20-40 words)" + letting the section 4 examples teach the style. No structural phrase = no template to copy.

New issues (different from template lock):

• llama3.1 consequence parroting in later calls (review ≈ consequence with modal verb swap). Call-2+ specific.
• haiku fabricated numbers +8.3pp (22.6%→30.9%) — likely from section 5 change, not the review_lever tweak.

This is the strongest result across all 8 iterations. The template lock was the #1 issue from analysis 40 and it's resolved.

[neoneye]

Closing — template lock fixed but haiku fabricated numbers regressed +8.3pp and llama3.1 consequence parroting in later calls. Not convincingly better overall.