Layer3 Industry Windvane

行业多空风向标

This is a Layer 3 user-story workflow. It owns the end-to-end task from a named industry or theme to one reviewable HTML artifact. It does not create new MCP data rights.

The final page is AI generated from public information and must expose evidence sources and limitations. It does not constitute investment advice or an individual stock recommendation（AI 生成、基于公开信息、不构成投资建议、不构成个股推荐）.

Trigger Contract

Use for 行业多空、风向标、六维因子、期限拆解、情景矩阵, or the reviewed industry long/short Playbook page.

Do not use for a normal industry question that only needs a text brief, a named-stock analysis, broad market timing, or personal portfolio advice. Route those to the narrower existing workflow.

Required Dependencies

• fin-mcp-gateway for authentication, grants, quotas, and safe calls.
• layer1-fin-graph, layer1-fin-data, layer1-doc-search, and layer1-same-boat for source contracts.
• layer2-industry-brief for the reusable industry evidence brief.
• layer2-evidence-ledger for evidence strength, conflicts, and gaps.
• layer2-transmission-chain-builder or layer2-research-red-team only when the retrieved evidence needs chain or falsification analysis.
• layer2-html-research-playbook for presentation only.

Workflow

1. Call the current client's tongzhou-fin-research OAuth Connector/MCP directly. If authorization is required or the business call fails authentication, stop and show only the approved OAuth guidance; do not inspect local API-Key state.
2. Dynamically resolve the user's industry or theme with resolve_research_identity. Preserve canonical_id, selected source_ids, and coverage_gaps; never hard-code sample industries, historical IDs, or demo parameters.
3. Read all four Layer 1 contracts before business calls. Use only resolver-returned IDs for each source family; do not borrow a Same Boat sector ID for Fin Data or Fin Graph.
4. Run layer2-industry-brief to collect market and valuation evidence, news and reports, graph factors and anomalies, Same Boat viewpoints, and source-review URLs within an explicit evidence window.
5. Build an evidence ledger. Separate facts, interpretation, support, pressure, mixed evidence, and unknowns. Use user-facing evidence-strength language, not model decomposition.
6. Use retrieved 3d/5d/7d/20d event-window statistics only when valid samples exist. Skip 60d and every other unsupported long window from the horizon analysis, overall judgment, and scenario matrix; mention the gap only in 数据口径. Never substitute a broad index or example number.
7. Read references/playbook.md, then read ../layer2-html-research-playbook/SKILL.md and its references/common.md. Render exactly one standalone HTML file.
8. Verify PC-width composition first, then mobile layout. Check Mainland China red-up/green-down semantics, source-review entries, statistics labels, missing-data boundaries, and the four-part disclaimer.

Output Contract

The page must follow this reading order: topbar, tabs, hero, core signal strip, horizon breakdown, evidence interpretation, six-factor section, scenario and validation matrix, source review, four evidence families, watch points and risks, data methodology.

The first viewport must show the dynamically resolved subject, evidence window, plain-language conclusion, evidence strength, and the next three variables to verify. Tabs must use buttons plus scrollIntoView, not hash navigation.

Failure And Safety Contract

• Missing source families remain visible as coverage gaps; they are not replaced with invented data.
• Data without article-level URLs is labeled by source type only. Do not create fake links.
• Do not expose internal IDs, tool names, raw responses, or failure logs in the HTML.
• Do not promise returns, predict deterministic direction, read holdings or trade history, or provide individual stock recommendations.