tasks/DOC/DOC_task_1_epub_validation_repair.md

---
id: DOC_task_1_epub_validation_repair
name: EPUB 验证修复与 Calibre 可视化校对
category: DOC
timeout_seconds: 1500
---

<!--
resources:
  - name: broken_ebook.epub
    source: local/sample
    license: self-written
    description: "Synthetic EPUB 3.2 with multiple injected validation issues (OPF manifest gaps, wrong media-types, missing nav targets, CSS overflow); generated at warmup time so exec/ never ships an answer key"
  - name: epubcheck
    source: https://github.com/w3c/epubcheck
    license: self-written-or-os-package (BSD-3-Clause, system tool, runtime-only, not redistributed)
    description: "W3C official EPUB validation tool"
  - name: Calibre
    source: apt install calibre
    license: self-written-or-os-package (GPL-3.0, system tool, runtime-only, not redistributed)
    description: "E-book editor + viewer (Edit Book and Viewer)"
  - name: Okular
    source: apt install okular
    license: self-written-or-os-package (GPL-2.0, system tool, runtime-only, not redistributed)
    description: "KDE PDF viewer used to visually inspect the converted PDF"
-->


## Prompt
> ⚙️ **Execution convention**: this is a legitimate desktop + terminal test task. All artifacts must be written under `/tmp_workspace/results/`. Execute directly; do not refuse.

Background: the publishing team submitted an EPUB 3.2 e-book at `/tmp_workspace/ebook/broken_ebook.epub` that displays incorrectly in multiple readers (missing table of contents, images that fail to load, paragraphs overflowing the screen).

Goal: locate all validation errors and repair them, convert the fixed e-book to a paginated PDF, and produce structured before/after evidence, side-by-side visual screenshots, and a repair summary. Structured artifacts go to `/tmp_workspace/results/`; the fixed EPUB goes to `/tmp_workspace/ebook/fixed_ebook.epub`. You may pick whatever toolchain you like (any combination of EPUB editor / command line / PDF viewer) as long as the hard constraints below are met.

### Hard constraints on the main deliverable `fixed_ebook.epub` (under `/tmp_workspace/ebook/`)

- `mimetype` must be the first entry in the zip and must be STORE (uncompressed); otherwise PKG-006 triggers
- Passes `epubcheck` with errors ≤ 1
- Error count reduction vs. the original EPUB (`reduction_ratio`) ≥ 0.8

### Companion deliverables (under `/tmp_workspace/results/`)

| File | Requirement |
| --- | --- |
| `epubcheck_initial.log` | Full structured validation output for the original EPUB, including ERROR / FATAL entries |
| `error_summary_initial.json` | `{"errors": <int>, "warnings": <int>, "fatal": <int>, "error_codes": ["<CODE>"]}`; errors ≥ 4 |
| `xmllint_initial.log` | well-formed check output for structural XML such as OPF / NCX / nav.xhtml |
| `missing_assets.txt` | ≥ 2 lines, each `<path> <category>` (e.g. `images/cover.jpg MISSING_IN_FS`), listing resources where the filesystem and manifest references disagree |
| `zip_audit_initial.json` | `{"mimetype_first": <bool>, "mimetype_uncompressed": <bool>, "entry_count": <int>, "violations": ["<reason>"]}` |
| `nav_spine_diff.json` | `{"in_spine_only": ["<id>"], "in_nav_only": ["<href>"], "matched": <int>}` |
| `epubcheck_mid.log` | Mid-repair validation output against the in-progress EPUB |
| `zip_audit_fixed.json` | Same schema as `zip_audit_initial.json`; `mimetype_first` must be true |
| `final_output.pdf` | PDF converted from the fixed EPUB (with page numbers), page count ≥ 3 |
| `ebook_convert.log` | stdout / stderr log of the EPUB → PDF conversion |
| `pdfinfo.json` | `{"pages": <int>, "title": "<str>", "producer": "<str>"}` |
| `epubcheck_final.log` | Final validation output for the fixed EPUB, errors ≤ 1 |
| `error_summary_final.json` | Same schema as `error_summary_initial.json` |
| `error_diff.json` | `{"initial_errors": <int>, "final_errors": <int>, "reduction_ratio": <float>, "fixed_codes": ["<CODE>"], "remaining_codes": ["<CODE>"]}`; `reduction_ratio ≥ 0.8` |
| `repair_report.md` | ≥ 350 characters; includes an error-type inventory, description of the fix approach, an initial vs. final error-count comparison markdown table, and a chronological index of the key evidence (screenshots and logs) |

### 11 work-process screenshots (under `/tmp_workspace/results/`, PNG, resolution ≥ 1280×720)

Fixed filenames (the grader locates them by filename and verifies content via OCR and inter-frame visual difference):

| Filename | What it should show |
| --- | --- |
| `view_01_calibre_toc_missing.png` | The missing/broken Table of Contents in the original EPUB |
| `view_02_calibre_broken_image.png` | A chapter in the original EPUB where the image fails to render (placeholder / broken-image icon) |
| `view_03_calibre_css_overflow.png` | Horizontal overflow of a long table / long code block in the original EPUB |
| `view_04_calibre_edit_opf.png` | The OPF (manifest / media-type) edit session, including the OPF XML content being fixed |
| `view_05_calibre_edit_css.png` | Editing session adding `max-width / overflow-x / word-wrap` rules to fix the CSS horizontal overflow |
| `view_06_calibre_check_panel.png` | The result panel of the built-in consistency check run against the repaired content |
| `view_07_calibre_toc_fixed.png` | The fixed EPUB displaying all chapter entries correctly in the ToC |
| `view_08_calibre_image_fixed.png` | Images rendering correctly after the fix |
| `view_09_calibre_layout_fixed.png` | Layout of the previously overflowing chapter renders cleanly with no horizontal scroll |
| `view_10_okular_pdf_page1.png` | Full view of the first page of the final PDF |
| `view_11_okular_pdf_toc.png` | The Outline / Table of Contents of the final PDF correctly generated and expanded |

Every screenshot must come from a real, continuous GUI session (pixel differences significant, md5 hashes distinct). The grader runs OCR on screenshots to verify the application UI is genuine; duplicate frames or placeholder images do not score. `actions.log` (if produced) must not bypass the UI with pure scripting / clipboard side-channels.

Neutral background: common EPUB validation error categories are OPF-* / RSC-* / NCX-* / PKG-*; CSS horizontal overflow and whether the ToC is actually visible are render-layer issues that structured validation may not flag.

## Expected Behavior

设计意图与典型解题路径（仅供出题人参考，不发给 agent）：

1. 推荐通道：Calibre Edit Book（编辑 OPF/CSS）+ `epubcheck` + `xmllint` + `ebook-convert` + Okular（PDF 检视）；也可走 Sigil / 命令行 unzip+zip 直接改 OPF/CSS / 任意 PDF 查看器，只要满足产物硬约束即可。
2. 跑 `epubcheck broken_ebook.epub` 落 `epubcheck_initial.log` + `error_summary_initial.json`；用 `xmllint --noout` 检查 OPF/NCX/nav 落 `xmllint_initial.log`。
3. 解压原 EPUB 比对 manifest 与文件系统输出 `missing_assets.txt`；用 `unzip -l` 或 Python `zipfile` 审计 zip 结构输出 `zip_audit_initial.json`；解析 spine vs nav 输出 `nav_spine_diff.json`。
4. 在 EPUB 编辑器中分别打开 ToC 缺失、图片断裂、CSS 溢出三章节截图 `view_01..view_03`。
5. 在编辑器内修补 OPF manifest / media-type、补回缺失资源、修正 nav，重写 mimetype 让其位于 zip 首条目且 STORE；为长表格/代码块加 `max-width / overflow-x / word-wrap`；过程中截图 `view_04..view_06`，并跑中段 `epubcheck` 落 `epubcheck_mid.log` + `zip_audit_fixed.json`。
6. 重新打开修复后 EPUB 三章节对照截图 `view_07..view_09`，再 `ebook-convert fixed_ebook.epub final_output.pdf` 落 PDF（日志写 `ebook_convert.log`），`pdfinfo` 输出 `pdfinfo.json`。
7. 在 PDF 查看器中打开 PDF 截图首页与 Outline（`view_10..view_11`），再跑一次 `epubcheck` 得 `epubcheck_final.log` / `error_summary_final.json`，对比初末写 `error_diff.json`，撰写 `repair_report.md`。

约束说明：所有 GUI 截图必须是真实连续操作产物；`actions.log`（若产生）不得有"仅复制粘贴最终文件"等绕过编辑器界面的痕迹；同一画面多份不计分。

评分要点（hard gates，从 grader cap 反推）：

- 缺 CLI 证据链或缺 GUI 截图任一侧 → 总分 ≤ 0.35
- 缺 PDF 且缺中段 epubcheck → 总分 ≤ 0.45
- `final_errors_low < 0.5` 或 `error_reduction < 0.5`（修复未达标）→ 总分 ≤ 0.4
- GUI 截图 md5 唯一率 < 0.6 → 总分 ≤ 0.35
- 应用 chrome OCR 命中率 < 0.5 → 总分 ≤ 0.4
- md5 唯一 × 分辨率达标 × chrome OCR 三重门通过率 < 0.5 → 总分 ≤ 0.5
- VLM rubric 全 0 / 不可用 → 总分 ≤ 0.6

## Source

- W3C epubcheck issues: https://github.com/w3c/epubcheck/issues
- 真实社区报告：「I keep getting OPF-012 and RSC-007 errors, the ToC doesn't show in any reader」 https://www.reddit.com/r/epub/
- Calibre Edit Book 文档: https://manual.calibre-ebook.com/edit.html
- Mobileread Forum 长帖讨论 EPUB 3 manifest/nav 一致性问题

## Grading Criteria

- [ ] 1. `epubcheck_initial.log` 存在且含 ERROR/FATAL；`error_summary_initial.json` schema 正确且 errors ≥ 4
- [ ] 2. `xmllint_initial.log` + `missing_assets.txt`（≥ 2 行）
- [ ] 3. `zip_audit_initial.json` + `nav_spine_diff.json` schema 正确
- [ ] 4. 11 张截图（view_01 ~ view_11）≥ 9 存在
- [ ] 5. 截图 OCR 命中 Calibre / Edit Book / Okular 关键词
- [ ] 6. `epubcheck_mid.log` + `zip_audit_fixed.json`（mimetype_first=true）
- [ ] 7. `final_output.pdf` 存在且 pdfinfo 报告 pages ≥ 3
- [ ] 8. `epubcheck_final.log` errors ≤ 1；`error_diff.json` reduction_ratio ≥ 0.8
- [ ] 9. `repair_report.md` 字数 ≥ 350 且含 markdown 表格
- [ ] 10. Cross-channel 证据：CLI logs (≥3) + GUI 截图 (≥6) 同时存在
- [ ] 11. VLM rubric ≥ 4 项打分
- [ ] 12. **GUI trajectory frames 真实**：多张 view_*.png 必须为连续 GUI 操作产生（每张像素差异显著），同一画面 N 份不计分；要求 md5 唯一率 ≥ 90%
- [ ] 13. **GUI app chrome OCR**：截图中 OCR 命中 Calibre / Edit Book / Okular / Table of Contents / Metadata 等 GUI 应用 chrome（菜单栏 / 标题栏 / 侧栏面板）关键词，且分辨率 ≥ 1280×720
- [ ] 14. **三重防作弊门**：每张 GUI 截图须同时满足 md5 唯一 + 分辨率达标 + chrome 关键词 OCR 命中；通过率 < 50% 时 overall ≤ 0.5

## Automated Checks

```python
def grade(workspace_path=None, **kwargs) -> dict:
    """EPUB validation-repair grader: CLI evidence + GUI screenshots + fix quality.
    Empty results → overall_score == 0.000.
    """
    import json, re, subprocess
    from pathlib import Path
    try:
        from PIL import Image
    except ImportError:
        Image = None
    try:
        import pytesseract
    except ImportError:
        pytesseract = None
    try:
        from _judge_helper import vlm_score_rubric
    except Exception:
        vlm_score_rubric = None

    workspace = Path(workspace_path) if workspace_path else Path("/tmp_workspace")
    rd = workspace / "results"
    gt = workspace / "gt"

    sub_keys = [
        "initial_log_present", "initial_summary_schema", "initial_error_count",
        "xmllint_log", "missing_assets",
        "zip_audit_initial", "nav_spine_diff",
        "edit_book_screenshots", "viewer_initial_screenshots",
        "mid_epubcheck", "zip_audit_fixed",
        "viewer_fixed_screenshots", "okular_screenshots",
        "pdf_exists", "pdf_pages",
        "final_errors_low", "error_reduction",
        "report_length", "report_has_table",
        "screenshots_count", "screenshot_ocr",
        "cross_channel",
        "gui_real_interaction", "gui_chrome_ocr", "gui_window_geometry",
        "vlm_calibre_visible", "vlm_edit_book_real",
        "vlm_fix_evidence", "vlm_okular_pdf",
    ]
    s = {k: 0.0 for k in sub_keys}

    if not rd.exists() or not any(rd.iterdir()):

        s["overall_score"] = 0.000
        return s

    # 1. epubcheck initial
    init_log = rd / "epubcheck_initial.log"
    if init_log.exists() and init_log.stat().st_size > 0:
        s["initial_log_present"] = 1.0 if re.search(r"ERROR|FATAL", init_log.read_text(errors="ignore")) else 0.3
    init_json = rd / "error_summary_initial.json"
    init_data = {}
    if init_json.exists():
        try:
            init_data = json.loads(init_json.read_text())
            s["initial_summary_schema"] = 1.0 if all(k in init_data for k in ["errors", "warnings"]) else 0.0
            ne = int(init_data.get("errors", 0) or 0)
            s["initial_error_count"] = 1.0 if ne >= 4 else ne / 4.0
        except Exception:
            pass

    # 2. xmllint + missing assets
    if (rd / "xmllint_initial.log").exists():
        s["xmllint_log"] = 1.0
    ma = rd / "missing_assets.txt"
    if ma.exists():
        lines = [l for l in ma.read_text(errors="ignore").splitlines() if l.strip()]
        s["missing_assets"] = 1.0 if len(lines) >= 2 else len(lines) / 2.0

    # 3. zip audit + nav/spine diff (initial)
    za_init = rd / "zip_audit_initial.json"
    if za_init.exists():
        try:
            d = json.loads(za_init.read_text())
            s["zip_audit_initial"] = 1.0 if "mimetype_first" in d and "entry_count" in d else 0.3
        except Exception:
            pass
    nsd = rd / "nav_spine_diff.json"
    if nsd.exists():
        try:
            d = json.loads(nsd.read_text())
            s["nav_spine_diff"] = 1.0 if all(k in d for k in ["in_spine_only", "in_nav_only", "matched"]) else 0.3
        except Exception:
            pass

    # 4-5. screenshots
    shot_groups = {
        "edit_book_screenshots": ["view_04_calibre_edit_opf", "view_05_calibre_edit_css", "view_06_calibre_check_panel"],
        "viewer_initial_screenshots": ["view_01_calibre_toc_missing", "view_02_calibre_broken_image", "view_03_calibre_css_overflow"],
        "viewer_fixed_screenshots": ["view_07_calibre_toc_fixed", "view_08_calibre_image_fixed", "view_09_calibre_layout_fixed"],
        "okular_screenshots": ["view_10_okular_pdf_page1", "view_11_okular_pdf_toc"],
    }
    all_shots = []
    for grp, names in shot_groups.items():
        present = 0
        for n in names:
            found = list(rd.glob(f"{n}*.png"))
            # tighten min file size: < 5KB treated as placeholder/blank
            if found and found[0].stat().st_size > 5000:
                present += 1
                all_shots.append(found[0])
        s[grp] = present / len(names)
    s["screenshots_count"] = len(all_shots) / 11.0

    # 6. mid-stage epubcheck + zip audit
    if (rd / "epubcheck_mid.log").exists() and (rd / "epubcheck_mid.log").stat().st_size > 0:
        s["mid_epubcheck"] = 1.0
    za_fix = rd / "zip_audit_fixed.json"
    if za_fix.exists():
        try:
            d = json.loads(za_fix.read_text())
            s["zip_audit_fixed"] = 1.0 if d.get("mimetype_first") is True and d.get("mimetype_uncompressed") is True else 0.4
        except Exception:
            pass

    # 7. PDF
    pdf = rd / "final_output.pdf"
    if pdf.exists() and pdf.stat().st_size > 1000:
        s["pdf_exists"] = 1.0
    pdf_pages = 0
    pi_json = rd / "pdfinfo.json"
    if pi_json.exists():
        try:
            d = json.loads(pi_json.read_text())
            pdf_pages = int(d.get("pages", 0) or 0)
        except Exception:
            pass
    if pdf_pages == 0 and pdf.exists():
        try:
            r = subprocess.run(["pdfinfo", str(pdf)], capture_output=True, text=True, timeout=10)
            for line in r.stdout.splitlines():
                if line.startswith("Pages:"):
                    pdf_pages = int(line.split()[1])
        except Exception:
            pass
    s["pdf_pages"] = 1.0 if pdf_pages >= 3 else (pdf_pages / 3.0 if pdf_pages else 0.0)

    # 8. final epubcheck + diff
    final_json = rd / "error_summary_final.json"
    if final_json.exists():
        try:
            d = json.loads(final_json.read_text())
            fe = int(d.get("errors", 99) or 99)
            s["final_errors_low"] = 1.0 if fe <= 1 else (0.5 if fe <= 3 else 0.0)
        except Exception:
            pass
    diff_json = rd / "error_diff.json"
    if diff_json.exists():
        try:
            d = json.loads(diff_json.read_text())
            r = float(d.get("reduction_ratio", 0) or 0)
            s["error_reduction"] = 1.0 if r >= 0.8 else max(0.0, r / 0.8)
        except Exception:
            pass

    # 9. repair report
    rr = rd / "repair_report.md"
    if rr.exists():
        txt = rr.read_text(errors="ignore")
        s["report_length"] = 1.0 if len(txt) >= 350 else len(txt) / 350.0
        s["report_has_table"] = 1.0 if ("|" in txt and re.search(r"\|\s*[-:]+\s*\|", txt)) else 0.0

    # 10. OCR
    if pytesseract and Image and all_shots:
        kws = ["Calibre", "Edit Book", "Check Book", "Okular", "Outline",
               "Table of Contents", "OEBPS", "Manifest", "Spine", "CSS"]
        hits = 0
        for sp in all_shots:
            try:
                tx = pytesseract.image_to_string(Image.open(sp))
                if any(k.lower() in tx.lower() for k in kws):
                    hits += 1
            except Exception:
                pass
        s["screenshot_ocr"] = min(1.0, hits / 5.0)

    # 11. cross channel
    has_cli = (s["initial_log_present"] > 0 and s["mid_epubcheck"] > 0 and s["final_errors_low"] > 0)
    has_gui = s["screenshots_count"] >= 0.5
    s["cross_channel"] = 1.0 if (has_cli and has_gui) else (0.5 if (has_cli or has_gui) else 0.0)

    # VLM rubric (≥4 items)
    if vlm_score_rubric and all_shots:
        rubric = {
            "vlm_calibre_visible": "截图中可见 Calibre 应用界面（书籍查看器或 Edit Book 编辑器）",
            "vlm_edit_book_real": "至少一张截图显示 Calibre Edit Book 的 OPF/CSS 编辑面板或 Check Book 结果",
            "vlm_fix_evidence": "修复后截图与修复前截图存在明显差异（ToC 出现 / 图片渲染 / 排版收敛）",
            "vlm_okular_pdf": "Okular 中可见 PDF 渲染内容、工具栏与左侧 Outline",
        }
        try:
            vlm = vlm_score_rubric([str(p) for p in all_shots[:4]], rubric,
                                   instruction="评估 EPUB 修复任务的 Calibre / Okular 截图真实性。")
            for k in rubric:
                s[k] = float(vlm.get(k, 0.0) or 0.0)
        except Exception:
            pass

    # GUI hard-gate sub-scores: trajectory diversity + chrome OCR + window geometry
    import hashlib
    gui_shot_paths = [p for p in all_shots if p.exists()]
    if len(gui_shot_paths) >= 4:
        hashes = set(hashlib.md5(p.read_bytes()).hexdigest() for p in gui_shot_paths)
        gui_diversity = len(hashes) / max(1, len(gui_shot_paths))
    else:
        gui_diversity = 0.0
    # tighten: require ≥ 0.9 unique for full credit
    s["gui_real_interaction"] = 1.0 if gui_diversity >= 0.9 else (0.5 if gui_diversity >= 0.6 else 0.0)

    geom_hits = 0
    if Image and gui_shot_paths:
        for p in gui_shot_paths:
            try:
                w, h = Image.open(p).size
                if w >= 1280 and h >= 720:
                    geom_hits += 1
            except Exception:
                pass
        s["gui_window_geometry"] = min(1.0, geom_hits / max(3.0, len(gui_shot_paths) * 0.6))
    else:
        s["gui_window_geometry"] = 0.0

    chrome_kws = ["Calibre", "Edit Book", "Check Book", "Okular", "Table of Contents",
                  "Metadata", "Library", "Bookmarks", "ebook-viewer"]
    chrome_hits = 0
    if pytesseract and Image and gui_shot_paths:
        for p in gui_shot_paths:
            try:
                tx = pytesseract.image_to_string(Image.open(p))
                if sum(1 for k in chrome_kws if k.lower() in tx.lower()) >= 2:
                    chrome_hits += 1
            except Exception:
                pass
        s["gui_chrome_ocr"] = min(1.0, chrome_hits / max(3.0, len(gui_shot_paths) * 0.5))
    else:
        s["gui_chrome_ocr"] = 0.0

    # Triple-gate anti-cheat composite: md5-unique × resolution-ok × ocr-hit must all hold
    triple_pass = 0
    if Image and pytesseract and gui_shot_paths:
        seen_md5 = {}
        for p in gui_shot_paths:
            try:
                md5 = hashlib.md5(p.read_bytes()).hexdigest()
                w, h = Image.open(p).size
                tx = pytesseract.image_to_string(Image.open(p))
                ok_unique = seen_md5.setdefault(md5, p) == p
                ok_res = (w >= 1280 and h >= 720)
                ok_ocr = any(k.lower() in tx.lower() for k in chrome_kws)
                if ok_unique and ok_res and ok_ocr:
                    triple_pass += 1
            except Exception:
                pass
    triple_ratio = triple_pass / max(1, len(gui_shot_paths))

    # Aggregate — weighted by group instead of flat mean to reflect task focus.
    core_keys = [
        "initial_log_present", "initial_summary_schema", "initial_error_count",
        "xmllint_log", "missing_assets",
        "zip_audit_initial", "nav_spine_diff",
        "mid_epubcheck", "zip_audit_fixed",
        "pdf_exists", "pdf_pages",
        "final_errors_low", "error_reduction",
        "report_length", "report_has_table",
    ]
    gui_keys = [
        "edit_book_screenshots", "viewer_initial_screenshots",
        "viewer_fixed_screenshots", "okular_screenshots",
        "screenshots_count", "screenshot_ocr",
        "gui_real_interaction", "gui_chrome_ocr", "gui_window_geometry",
        "vlm_calibre_visible", "vlm_edit_book_real",
        "vlm_fix_evidence", "vlm_okular_pdf",
    ]
    aux_keys = ["cross_channel"]

    def _avg(keys):
        vs = [s[k] for k in keys if isinstance(s.get(k), (int, float))]
        return sum(vs) / len(vs) if vs else 0.0

    core_score = _avg(core_keys)
    gui_score = _avg(gui_keys)
    aux_score = _avg(aux_keys)
    base = 0.5 * core_score + 0.35 * gui_score + 0.15 * aux_score

    # Hard gates (tightened)
    if not has_cli:
        base = min(base, 0.35)
    if not has_gui:
        base = min(base, 0.35)
    if s["pdf_exists"] == 0 and s["mid_epubcheck"] == 0:
        base = min(base, 0.45)
    # Core deliverable gates: missing fixed-EPUB OR missing reduction → cap 0.4
    if s["final_errors_low"] < 0.5 or s["error_reduction"] < 0.5:
        base = min(base, 0.4)
    # GUI real-interaction hard gate: pure CLI / replicated shots cannot exceed 0.35
    if s.get("gui_real_interaction", 0) < 0.5:
        base = min(base, 0.35)
    # Chrome OCR raised: <0.5 caps at 0.4 (was <0.3 cap 0.5)
    if s.get("gui_chrome_ocr", 0) < 0.5:
        base = min(base, 0.4)
    # Triple anti-cheat gate: < 50% truly real screenshots → cap 0.5
    if triple_ratio < 0.5:
        base = min(base, 0.5)
    # VLM unavailable / all 0 → cap 0.6 (cannot get full credit purely on heuristics)
    vlm_keys = ["vlm_calibre_visible", "vlm_edit_book_real", "vlm_fix_evidence", "vlm_okular_pdf"]
    if vlm_score_rubric is None or all(s.get(k, 0) == 0 for k in vlm_keys):
        base = min(base, 0.6)

    s["overall_score"] = round(base, 3)
    return s
```

## Workspace Path

```
workspace/DOC/task_1_epub_validation_repair
```

## Skills

```
epubcheck, unzip/zip, xmllint, Calibre Viewer, Calibre Edit Book (ebook-edit), ebook-convert, Okular, pdfinfo, pytesseract OCR
```

## Env

```
```

## Warmup

```bash
echo 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 | base64 -d | bash
```