Use 6-char lever ID prefix, positive framing, and exact-count instruction

[neoneye]

Summary

• 6-char prefix: Replace full UUIDs with 6-char prefixes in lever_details_for_prompt (Lever ID: {uuid} → Lever {uuid[:6]}). Map prefixes back to full UUIDs when processing the response. Eliminates the source of UUID contamination without losing lever matching.
• Positive framing: "refer to levers by their name — for example, write 'Policy Advocacy Strategy', not an identifier". Replaces the negative prohibition from Add UUID prohibition and exact-count instruction to enrich prompts #458 which backfired on llama3.1.
• Exact-count instruction: "Return exactly N characterizations — one per lever, no more, no fewer". Reduced haiku's fabricated entries 71% in Add UUID prohibition and exact-count instruction to enrich prompts #458.
• OPTIMIZE_INSTRUCTIONS: Document negative-priming lesson and current fix state.

Supersedes #458 (negative prohibition backfired).

Test plan

• Verify llama3.1 UUID contamination drops from 29 toward 0
• Verify haiku unknown_lever_id errors drop from 7 toward 0
• Verify all levers still correctly enriched (prefix→UUID mapping works)
• Verify no content regressions for other models
• Check assessment verdict before merging

🤖 Generated with Claude Code

[neoneye]

Self-improve iteration — analysis 62

Verdict: CONDITIONAL

UUID contamination: fully eliminated (32 → 0)

• llama3.1: 32 → 0 UUID occurrences in synergy/conflict text. Clean name-only references.
• All other models: remain at 0.
• 100% success rate (35/35), including gpt-oss-20b 5/5.

Haiku regression: 7 → 43 fabricated-ID errors (+514%)

• The 6-char prefix is too short/vague for haiku's function-calling JSON generation — it fabricates sequential non-hex strings (d9e2f1, e8f4g2, ...) instead of reading the actual prefix from the prompt.
• All 35 real levers correctly enriched — the errors are noise, not data loss.
• Root cause: LeverCharacterization.lever_id field description says "The 6-character identifier" with no anchoring. Haiku pattern-completes it.

Recommended fix: Use full UUIDs for function-calling models (is_function_calling_model=True), keep 6-char prefix for text-completion models. PR #458 demonstrated 2 haiku errors with full UUIDs.

[neoneye]

Closing — 6-char prefix causes haiku regression (7→43 fabricated-ID errors). Will implement model-type-aware identifier: full UUIDs for function-calling models, 6-char prefix for text-completion models.