Replace negative-priming prohibition with positive framing in consequences field

[neoneye]

Summary

Replace "Do NOT include 'Controls ... vs.', 'Weakness:'" with "Focus on cause-effect relationships and factual outcomes; save critical assessments for the review_lever field." in the consequences field description of both Lever and LeverCleaned Pydantic models.

The negative prohibition names exact banned phrases that small models copy as templates — the same anti-pattern that caused regressions in the enrich step (PRs #458, #460). Positive framing guides models toward desired behavior without priming the unwanted pattern.

Test plan

• Run self_improve experiment for identify_potential_levers step
• Verify llama3.1 consequences field no longer contains "Controls ... vs." or "Weakness:" patterns
• Verify no regression in consequences quality for other models
• Check assessment verdict before merging

🤖 Generated with Claude Code

[neoneye]

Self-improve iteration — analysis 69

Verdict: YES

• Positive framing correctly replaces negative prohibition in both Lever and LeverCleaned models
• "Controls X vs Y" / "Weakness:" patterns not observed in any after-batch outputs
• No content regressions — consequences field produces clean cause-effect prose
• Success rate: 97.8% (excl. gpt-oss-20b infrastructure timeout, not PR-caused)
• OPTIMIZE_INSTRUCTIONS alignment improved (lines 80-83 no longer violated)

[neoneye]

Closing — gpt-oss-20b timed out on all 5 plans (infrastructure issue needs investigation before merging).