Qualitative consequences, positive framing, and consistent length targets

[neoneye]

Summary

• Qualitative consequences: Replace all number-evidence constraints with "Exact numbers will be determined further downstream — here the goal is to articulate the cause-effect relationship clearly." Section 5: "NO specific numbers, percentages, or monetary amounts — describe effects qualitatively."
• Positive framing: Replace "Do NOT include 'Controls ... vs.'" (same as prior PRs)
• Consistent length targets: "2-3 sentences" in all three sources (field desc ×2, system prompt). review_lever "One sentence (20-40 words)" matches system prompt section 6.

PR #475's "Never invent" constraint made fabrication 4x worse — models interpreted it as a permission floor and derived numbers by arithmetic from plan figures. This PR removes the number discussion entirely.

Test plan

• Verify fabricated numbers reduced (HKD refs: was 20 in Positive framing, consistent word counts, and stronger number-evidence constraint #475, 5 in baseline)
• Verify gpt-oss-20b completes 5/5 plans
• Verify haiku consequences ≤ 2.0x baseline
• Verify no content quality regression
• Check assessment verdict before merging

🤖 Generated with Claude Code

[neoneye]

Self-improve iteration — analysis 72

Verdict: CONDITIONAL

Wins:

• Fabricated numbers: -48% overall, haiku -70% (20→6)
• Review template lock ("All three options...") appears resolved (85%→0%)
• gpt-oss-20b 5/5 success (gpt-oss-20b parasomnia timeout is infrastructure noise)
• No content quality regressions on successful plans

Residual:

• 3 plan timeouts (gpt-oss-20b parasomnia, gpt-5-nano sovereign_identity/silo) — likely infrastructure noise
• qwen3-30b unchanged at 3 pct claims — "Exact numbers will be determined further downstream" may anchor it
• 6 remaining haiku pct claims in engineering-domain plans

Conditions: Confirm timeouts are noise (re-run), and consider removing the "downstream" anchor phrase.

[neoneye]

Closing — not convincingly better than baseline. Will refine: keep stricter review_lever and shorter consequences, but allow plan-provided numbers instead of banning all numbers outright.