fix: anonymize few-shot examples in identify_documents prompts

[neoneye]

Summary

• IDENTIFY_DOCUMENTS_*_SYSTEM_PROMPT variants in worker_plan_internal/document/identify_documents.py contained concrete, domain-named examples (childcare subsidy policies, housing price indices, mental health survey data, marathon training, climate crop yield, etc.) that the LLM reproduced verbatim for unrelated plans.
• Observed in the Denmark euro-adoption run (PlanExe-web/20260129_euro_adoption/): identified_documents_to_find.json was populated with the exact social-policy document names from the business prompt's NAMING CONVENTION and EXAMPLE MAPPING sections. Downstream DraftDocumentsToFindTask (separate per-document LLM call) correctly filled euro-adoption content into essential_information, producing a chimera output in draft_documents_to_find.json.
• Fix: replace concrete named examples in the Documents-to-Find sections (EXAMPLE MAPPING, FORBIDDEN, NAMING CONVENTION) with bracketed placeholders across all three variants (business / personal / other), and mark the EXAMPLE MAPPING blocks as illustrative patterns only — do NOT reuse these topics. Structural guidance is preserved; only the domain hooks the LLM was latching onto are removed.

Root cause

The prompt's illustrative examples acted as a domain prior. For a Denmark euro-adoption plan, the LLM generated documents named Existing National Childcare Subsidy Policies, Data on Average Childcare Costs, Tax Code Sections Related to Dependents, National Housing Price Indices, Official National Mental Health Survey Data, Existing Social Support Program Details — all literal strings from the system prompt examples.

Test plan

• Re-run the Denmark euro-adoption plan and confirm identified_documents_to_find.json contains euro-adoption-relevant source materials (treaty texts, ERM II convergence criteria, ECB monetary policy, payment-system standards, currency-conversion studies), not social-policy items.
• Spot-check runs on an unrelated domain (e.g., infrastructure, tech rollout) to confirm the prompt still yields well-named documents without the concrete examples.
• Consider queuing this prompt in the self_improve loop — it has not been iterated on yet.

🤖 Generated with Claude Code