Five-Layer Sieve Investment Analysis

Problem Addressed

• Single-stock research often mixes data, fundamentals, valuation, and decisions, which can reward first-impression bias and silently force contradictory findings into one conclusion.
• For engineers, the core issue is not only accuracy but auditability: which inputs came from scripts, which came from search, and where one layer contradicts an earlier one.

How It Works

• Multi-head attention style orchestration: maps the ticker into Query, layer rules into Key, and layer conclusions into Value; Head 0 through Head 5 cover data validation, positioning, drivers, conflicts, valuation, and decisions.
• Script-first evidence: runs existing scripts before falling back to web_search, reducing the chance that second-hand descriptions are treated as primary data.
• Cross-layer consistency checks: after each Head, the skill checks whether its conclusion conflicts with earlier outputs; if conflict exists, it pauses for user judgment rather than auto-reconciling.
• Higher-order synthesis: full reports continue into Head 6 to produce a more advanced forecast and a structured Markdown output.

Boundaries

• Best suited for investment-analysis orchestration on tickers with clear financial data and industry context, not for trade execution or real-time quotes.
• Real-time pricing, compliance opinions, and transaction decisions still need external systems and human review.