Phase 1   Data Collection
  └─ Automatic chaining at the end of fetch_data.py:
       build_inventory.py → data_inventory.md
       scan_anomalies.py  → anomalies.{json,md}

Phase 1.5 Data Checkpoint
  └─ Agent must read data_inventory.md + anomalies.md
  └─ ❌ Missing fields must be supplemented via web_fetch or explicitly marked [GAP]
  └─ CRITICAL anomalies must be fully discussed in Step 0.5

Phase 2  In-depth Analysis (Three-section Check System + Tags)
  §1 (Step 0, 0.5, 1, 2) → part1 → verify_facts.py --partial part1
  §2 (Step 3, 4, 5)       → part2 → verify_facts.py --partial part2
  §3 (Step 6, 7, 8)       → part3 → verify_facts.py --partial part3
  Merge: cat part*.md > report.md

Phase 2.5 Fact-checker
  └─ verify_facts.py FULL → Must have 0 FAIL
  └─ Task(general-purpose subagent) → fact-check-report.md
  └─ FAIL items must be edited and re-run (max 3 rounds)

Phase 2.6 Bear-case
  └─ Task(general-purpose subagent) → bear-case.md
  └─ Append to report.md as a new section

Phase 3   HTML Writing in 6 Batches
  └─ verify_html.sh (includes verify_content.py) → 0 FAIL
  └─ Delete intermediate products _h_part*.html

You are a reviewer. Check the numerical accuracy and inference rationality of report.md.

【Input】report.md / data.json / anomalies.md (path in current output directory)

【Tasks】
1. Sample 20 [F:] tags, verify the accuracy, unit and context of each value one by one
2. Check if the basis for all [I:] inferences is reasonable (broad reasons like "industry common sense · leading position" are not acceptable)
3. Check if the multiplication table for all [T:] target prices is self-consistent
4. Identify anomalies in anomalies.md that are not fully discussed in the report
5. Find internal contradictions in the report

【Output】fact-check-report.md
- PASS: N / FAIL: M / WARN: K
- Each FAIL item must include: location + problem description + repair instruction

You are a bearish analyst. Based on report.md and anomalies.md, write a 300-500 word opposing view.

【Tasks】
1. Find 3 pieces of negative evidence that the main agent did not fully discuss
2. Provide specific paths for "if the bullish view is wrong, where is it wrong"
3. Refute key [I:] inferences in the report
4. Propose 3 observable indicators: "if signal X appears, the bearish view is confirmed"

【Output】bear-case.md (no tags required)
- Must cite at least 2 IDs from anomalies.md

output/
└── <Stock Name>_<Code>/
    ├── <YYYY-MM-DD>/
    │   ├── data.json          # Phase 1
    │   ├── report.md.part1-3  # Intermediate products of Phase 2
    │   ├── report.md          # Merged final version of Phase 2
    │   ├── _h_part0-5.html    # Intermediate products of Phase 3
    │   └── report.html        # Final deliverable
    └── latest -> <YYYY-MM-DD>  # Symbolic link