每一步都留下 可检查的输出

{
  "citation_mapping": {
    "[1]": {
      "old_key": "Smith2020",
      "new_key": "Smith2020_JMLR",
      "status": "resolved",
      "source": "word_footnote"
    },
    "[2]": {
      "old_key": null,
      "new_key": "Zhang2021_NeurIPS",
      "status": "imported",
      "source": "zotero_export"
    }
  },
  "stats": {
    "total_citations": 47,
    "resolved": 45,
    "ambiguous": 2
  }
}

{
  "check_id": "check_references",
  "timestamp": "2026-04-28T09:14:32",
  "summary": {
    "total_refs": 83,
    "broken_citekeys": ["Wang2022_unknown"],
    "duplicate_labels": [],
    "unused_bibentries": ["Li2019_old"]
  },
  "severity": "WARN",
  "items": [
    {
      "type": "broken_citekey",
      "citekey": "Wang2022_unknown",
      "location": "chapter3.tex:142",
      "message": "cite key not found in .bib"
    }
  ]
}

--- a/chapter3.tex
+++ b/chapter3.tex
@@ -142,7 +142,7 @@
-As shown in \cite{Wang2022_unknown}, the approach
+As shown in \cite{Wang2022_CVPR}, the approach
 achieves competitive results.

-## dry-run summary
+## preview only — no changes applied
   files_would_change: 1
   replacements: 1
   mode: dry-run (pass --apply to execute)

{
  "verdict": "WARN",
  "dimensions": {
    "references":    { "status": "PASS", "reason": "all cite keys resolve" },
    "language":      { "status": "WARN", "reason": "2 style warnings remain" },
    "format":        { "status": "PASS", "reason": "labels and refs are stable" },
    "compile":       { "status": "WARN", "reason": "skipped in demo mode" },
    "export":        { "status": "WARN", "reason": "not produced by smoke test" },
    "review_loop":   { "status": "WARN", "reason": "not produced by smoke test" }
  },
  "next_actions": [
    "Review reports/check_language-report.json",
    "Re-run without --skip-compile before final submission"
  ]
}

{
  "review_todos": [
    {
      "id": "R3-1",
      "location": "chapter2.tex:89",
      "original": "The method is straightforward.",
      "advisor_comment": "too vague — explain the trick",
      "priority": "high",
      "status": "open"
    },
    {
      "id": "R3-2",
      "location": "fig4_caption",
      "original": "Figure 4: Results.",
      "advisor_comment": "add error bars",
      "priority": "medium",
      "status": "open"
    }
  ],
  "stats": { "total": 7, "high": 2, "medium": 3, "low": 2 }
}

## Defense Outline

### 1. Introduction (3 min)
- Motivation: real-time demand forecasting gap
- Contribution: unified graph + temporal model

### 2. Method (8 min)
- 2.1 Graph construction (slide 4-6)
- 2.2 Temporal encoder (slide 7-9)
- Highlight: adjacency learned jointly, not pre-defined

### 3. Experiments (5 min)
- Dataset: NYC-Taxi, PeMS-BAY
- Key figure: Fig.3 — accuracy vs. latency tradeoff

### 4. Conclusion (2 min)
- Takeaway: 15% MAE reduction at 30% lower latency