literaturereport

name	literaturereport
description	Use when converting a local English science, engineering, or medical paper PDF into a Chinese literature-report PPTX with source-grounded figures, directly readable speaker notes, Qi-2022-based layouts, or an optional user visual theme.

Literature Report

Non-negotiable contract

Use local Windows CPU processing as the primary path for PDF parsing, source mapping, Figure extraction and recovery, panel handling, PPTX compilation, rendering, and QA. The only pre-evidence-packet model exception is the AI-vision panel-label coordinate workflow defined in references/figure-splitting.md: after resolving every main Figure, use exactly one batch_panel_label_coordinates request for all Figures whose legends declare at least four panels. Give the model one contact sheet followed by each dense Figure as an independent original-resolution input in paper source order. Outside that exception, use model reasoning only after a pipeline-generated evidence packet exists. early-evidence-packet.json is a provisional packet for non-result slides only; result slides and final validation must use the final evidence_packet.json. Model reasoning remains limited to evidence-grounded narrative selection, Chinese slide copy, and complete Chinese speaker notes.

Never trade these requirements for speed:

• Do not reason directly from an unparsed PDF or skip the evidence packet.
• Never truncate a panel; preserve the complete Figure when no safe horizontal split exists.
• Do not replace directly readable speaker notes with prompts or presenter cues.

Inputs and style

• Require one local, selectable-text, English PDF. Do not accept a URL.
• Support exactly two paper types: original research and ordinary narrative review. Immediately after local text extraction, classify the PDF from its title, abstract language, and section structure before selecting a narrative.
• Accept an optional user PPTX as a visual-style source only.
• Without a user PPTX, use assets/Qi-2022.pptx as both the default style and page framework and read the bundled assets/Qi-2022.style-tokens.json; do not rescan the unchanged default deck. Cache tokens for an optional user style by the style deck, Qi deck, and inspector-code fingerprints.
• With a user PPTX, extract its fonts, colors, gradients, recurring header/footer logos, and visual treatments, then retain the Qi-2022 page framework. Never import the user's slide coordinates, order, or content.
• Read references/qa-rules.md before style extraction or final QA.

Cover slide

• The first slide must use the cover narrative role and Qi-2022 slide 1.
• Populate the root cover object defined in references/deck-contract.md.
• Use the exact English paper title, centered Times New Roman 42 pt bold in #064A91.
• Write one concise Chinese summary line, centered 32 pt bold in #1F2937. Use Microsoft YaHei for CJK text and Times New Roman for Latin letters, numbers, and scientific tokens.
• Compose the journal line as {journal}, {publication year}, IF={latest JIF} in centered 24 pt black. Use Microsoft YaHei for CJK text and Times New Roman for English, numbers, and scientific tokens.
• Keep at least 0.33 inches of clear vertical space between the Chinese summary box and the journal box. Preserve the journal, presenter, and date as one aligned metadata group when increasing this spacing.
• For every deck, verify the latest JCR Journal Impact Factor from an official current journal or publisher metrics page and store the value, source URL, and lookup date in the cover object. Do not infer or reuse a historical value from the paper PDF.
• Render 汇报人：{presenter} and the compilation-date line in centered 24 pt gray. Use Microsoft YaHei for CJK text and Times New Roman for English and numbers. The compiler supplies the report date automatically.
• Lower the inherited Qi-2022 English title box by 0.36 inches while preserving its width and height.
• Disable cover-text auto-shrink so the authored point sizes remain the visible point sizes in the rendered PPTX.
• Without a user template, use #064A91 for the English paper title. With a user template, replace only that title color within the cover with the extracted colors.primary; keep the fixed cover structure, fonts, sizes, and other cover colors unchanged.
• Do not show source labels, page numbers, logos, DOI, affiliations, or extra placeholders on the cover.

Agenda slide

• Every deck must use Qi-2022 slide 2 as an agenda immediately after the cover.
• Keep the fixed left gradient block and six numbered rows. Do not copy agenda coordinates or content from a user template.
• Use Microsoft YaHei 54 pt bold white for 目录, Times New Roman 36 pt bold #064A91 for 01 through 06, and Microsoft YaHei 28 pt bold #064A91 for the six Chinese labels. Disable text auto-shrink.
• Right-align the number column, left-align every Chinese label to one shared left edge, and bottom-align each number-label pair within its row.
• Without a user template, preserve the inherited Qi blue gradient on the left rounded rectangle. With a user template, use colors.primary as the dark stop and a 60% white tint of that color as the light stop; keep the gradient vertical and keep 目录 white.
• Original research uses exactly: 基本信息, 研究背景, 研究思路, 研究结果, 结论与讨论, 创新点与启发.
• Narrative review uses exactly: 基本信息, 综述背景与范围, 领域知识框架, 主题证据与进展, 共识、争议与不足, 研究空白与展望.

Basic-information slide

• Every deck must use Qi-2022 slide 3 as paper_info immediately after the agenda. Cover, agenda, and basic information are fixed slides for both original research and narrative reviews.
• Populate the root paper_info object defined in references/deck-contract.md.
• Automatically crop the visible journal logo, paper title, and author block from the first PDF page with scripts/extract_paper_header.py. Preserve the full page-header width and stop below the authors so the abstract and body are excluded. Do not ask the user to make this crop.
• Set the top title to 基本信息 in Microsoft YaHei 30 pt bold #064A91.
• In the left lower blue box, write 研究内容：{concise Chinese summary} and 关键词：{keywords}. Format only the fixed labels 研究内容： and 关键词： as Microsoft YaHei 16 pt bold. After each label, use Microsoft YaHei 16 pt regular for Chinese and Times New Roman 16 pt regular for English, numbers, and scientific tokens. Keep all text white and disable auto-shrink in this box.
• Limit paper_info.content_summary to 75 Unicode characters, excluding the compiler-added 研究内容： label. Within that limit, prioritize the study method, core mechanism, and human or clinical evidence rather than a generic conclusion.
• In the right rounded rectangle, write 第一通讯作者：{first corresponding author} in 16 pt bold #064A91; then write the PDF-derived author affiliation and 研究方向：待填写 in black 16 pt regular text.
• Below that, write the journal, JCR quartile, latest verified JIF, and exact publication date in black 16 pt regular text. Use Microsoft YaHei for Chinese and Times New Roman for English, numbers, and scientific tokens.
• Set 代表作 to Microsoft YaHei 18 pt bold #064A91 and place the fixed text 待填写 below it. Never browse for or auto-populate research direction or representative works.
• Allow auto-shrink only in the right-side author, affiliation, journal, and representative-work text boxes. Do not auto-shrink the title or left summary.

Content headers and fixed background slide

• Every content slide with a top section label, beginning with 基本信息, uses Microsoft YaHei 30 pt bold. Apply Times New Roman 30 pt bold to English and numeric runs. Preserve the inherited color and position and disable auto-shrink. Cover and agenda are excluded.
• Both research_background and review_background use Qi-2022 slide 4 with the fixed modules 当前挑战, 研究潜力, 现有研究不足, and 核心问题 in that order. Module titles are Microsoft YaHei 18 pt bold.
• Populate background_sections. Every module contains exactly three concise points, and every point starts with • . Body text uses Microsoft YaHei 16 pt regular with Times New Roman for English, numbers, and scientific tokens. Do not auto-shrink.
• Keep every complete visible background point, including its • marker, between 23 and 25 Unicode characters, targeting 25 characters and one rendered line. Enrich sparse wording or rewrite overflow; never shrink the 16 pt text. Use normalized single-line text: no leading, trailing, repeated, or newline whitespace may be used to pad the count, and each point must have at least 20 non-whitespace characters.
• Derive all four modules only from the paper Abstract and Introduction. For a narrative review without an explicit Introduction heading, an opening Background section may substitute. Do not use Results or Discussion claims as background evidence.
• Preserve background evidence anchors, the source_label, and complete speaker notes in the deck specification, but do not add a visible fourth source line inside any of the four cards.

Research-idea slide and workflow confirmation

• For every original paper, complete the one pdffigures2 attempt and the background-contrast recovery pass before requesting explicit workflow confirmation. Narrative reviews skip this workflow question.
• If every main Figure is available, create a workflow-only request. If any Figure remains user_crop_required, create one combined request containing workflow_request and every unresolved item under figure_requests. Do not ask a second workflow question after the Figure crops arrive.
• A present decision requires one user screenshot. Copy it unchanged into the reviewed Figure directory and register it as workflow_user_provided with extraction_status: user_provided. Do not crop, OCR, or analyze the workflow screenshot. An absent decision uses the no-workflow layout and has no workflow Figure.
• Do not use workflow_candidates to infer the answer or bypass user confirmation. That field is compatibility-only.
• Populate research_idea_content with one purpose, one research object, and exactly four functional method groups. For each new paper, derive all four method names and uses from that paper's Methods section. Group or merge the paper's actual techniques by experimental function when the source lists more or fewer than four; method labels shown in examples are never reusable defaults. Do not repeat method labels from a prior paper, and do not invent a filler method when the Methods evidence is absent.
• The page header is 30 pt bold with no auto-shrink. 研究目的 and 研究对象 use 20 pt bold; their explanations use 16 pt regular with auto-shrink. Bottom method names use 14 pt bold, and method uses use 14 pt regular; both may auto-shrink. Chinese uses Microsoft YaHei and Latin runs use Times New Roman.
• With a workflow, place only workflow_user_provided into inherited Qi slide 5 Picture 7 using contain fit. Without a workflow, replace that image with a compact 2x2 matrix headed 样本与对象, 数据来源, 分析层级, and 验证层级. Each module has two or three • points and does not imply sequence.
• Keep the lower method container top fixed and extend its height to 1.20 inches. Extend each method-use text box downward to 0.42 inches.
• Never place workflow_user_provided on any other slide. When a paper Figure render asset is itself a workflow or experimental-design schematic already represented by the confirmed user workflow, exclude that asset from result slides and explicitly select the next evidence-bearing Figure asset order.

Technical-route slide

• Every original-research deck uses Qi-2022 slide 6 as technical_route; narrative reviews retain their existing review framework.
• Populate technical_route_content with one evidence-grounded overall_route and exactly four ordered steps labelled 第一步 through 第四步. Each step contains a paper-specific overview and purpose; do not derive these columns from visible_points or reuse example-paper content.
• Set the page header to 技术路线 at 30 pt bold with no auto-shrink. Set the overall route to 20 pt bold with auto-shrink.
• Set step labels to 12 pt bold white, overviews to 17 pt bold, and purposes to 16 pt regular black. Use no auto-shrink for all three step columns.
• Limit every overview to 20 Unicode code points and one rendered line.
• Limit every purpose to 50 Unicode code points and at most two rendered lines. Rewrite overflowing text instead of reducing the point size.
• Use Microsoft YaHei for Chinese and Times New Roman for English, numbers, abbreviations, and scientific tokens.

Original-research result slides

• Use Qi-2022 slide 7 for every result or results slide. Each slide uses exactly one Figure render asset and records its exact one-based figure_asset_order; do not infer the asset only from slide occurrence.
• Populate visible_points from the displayed panels' Figure legend. When the asset contains more than two panel labels, write one to five bullets in the exact form • Fig.1B：结论 or • Fig.1C-D：合并结论. Each bullet may merge at most three adjacent panels; never collapse a long span such as Fig.2G-M into one item. Cover every displayed panel once. If exact coverage would require more than five bullets, stop before compilation with deck_validation_failed and report that an additional reviewed Figure split is required.
• Make every sidebar conclusion substantive: after ：, use at least 14 non-whitespace Unicode characters to state the comparison or readout together with the finding, while keeping the complete bullet at no more than 48 Unicode characters and all sidebar bullets together at no more than 155 characters. Store no whitespace directly after ： in deck_spec.json; the compiler uses a widened 17 pt text box, inserts one blank line between bullets, renders exactly one space after ：, and keeps the Figure range and colon together.
• When the displayed asset contains at most two panel labels, leave visible_points empty, remove the side text box, and expand the Figure across the full upper content area.
• Populate result_summary for every result slide in the exact form Fig.1B-E：总结. Derive this fixed bottom-box summary jointly from the Figure legend and the Results prose that discusses the displayed panels. Cover the full displayed panel range and keep the complete line at no more than 45 Unicode characters. Render the summary at 20 pt bold. Its full Figure range is intentionally exempt from the three-panels-per-sidebar-item limit.
• Place the side text box on the right for the first result slide that needs one. Alternate right, left, right, left across subsequent eligible result slides. A slide without a side text box does not advance the alternation.
• Keep Figure explanation and evidence anchors in deck_spec.json and speaker notes. Do not render source_label or any visible source box on the slide.

Conclusion-and-discussion slide

• Every original-research deck uses Qi-2022 slide 21 for conclusion_discussion and populates conclusion_discussion_content from the current paper's Results, Discussion, and Figure legends.
• If a Figure caption or legend explicitly identifies the Figure as a graphical abstract (case-insensitive), always treat its complete render asset as the independent mechanism Figure. This explicit label overrides other mechanism-role heuristics. Otherwise, require the Figure's primary purpose to be a final mechanism, working model, cross-talk model, or integrative schematic supported by the legend or Results text. A workflow diagram, ordinary result plot, or unsplit dense Figure containing only a small model panel does not qualify. Never generate a replacement mechanism image.
• With an independent mechanism Figure, set mechanism_figure_present: true, select exactly one explicit figure_id and figure_asset_order, place it in the inherited left image slot, keep the four-step evidence chain on the right, and render the integrated conclusion in the inherited bottom bar. Reserve that asset for this slide instead of duplicating it as a result page. A detected graphical abstract must therefore appear only on conclusion_discussion, never on a result or results slide.
• Without an independent mechanism Figure, set mechanism_figure_present: false, use no Figure ID or asset order, delete the inherited image, and expand the evidence-chain container and its text across the full content width. Never leave a blank image column.
• Author one evidence-chain heading, exactly four paper-specific labelled evidence steps, and one bottom integrated conclusion of at most 60 Unicode characters. The four labels describe the actual evidence chain and are not fixed to single-cell or spatial-omics examples. Keep visible_points empty.

Innovation and limitations slides

• Use exactly three content_blocks on innovation. Render every card heading, from the first innovation category through the clinical-connection category, in bold white. Render all three card bodies at Microsoft YaHei 18 pt with Times New Roman for Latin runs, use one uniform size across left, middle, and right cards, and disable auto-shrink.
• Use exactly four content_blocks on limitations_optimization. The first three blocks are three distinct evidence-grounded limitations and render under the fixed labels 局限1, 局限2, and 局限3; the fourth block renders under 优化方向 and proposes concrete responses to those limitations. Use each block title as a concise paper-specific summary such as 因果链不足 or 功能实验缺口; render these summaries in 18 pt bold #064A91 between the fixed label and the detailed explanation. Do not repeat the fixed label in the block title.
• Give every limitation or optimization body 30 to 70 non-whitespace Unicode characters. State the missing evidence or design constraint and its effect on interpretation; do not stop at a short phrase such as “样本量较小”. Render every black body at 18 pt, allow up to two lines, and disable auto-shrink.

Closing slide

• End every deck with exactly one closing slide using Qi-2022 slide 24 and template_frame: qi-closing.
• Set its visible title to 恳请各位老师批评指正, keep visible_points, figure_ids, source_label, and evidence_anchors empty, and preserve the inherited central rounded rectangle, underline, and bottom gradient bar.
• Add at least 60 Chinese characters of directly readable closing speaker notes; the notes may briefly recap the paper and invite questions, but the slide itself remains minimal.

Stateful local workflow

1. Create pipeline-config.json with the local PDF, work directory, pdffigures2 JAR, optional style PPTX, and installed Java and Codex Node paths. Check the runtime before the first pipeline run:

   python scripts/check_runtime.py --pdffigures-jar <jar> --java <java> --node <node>
   

2. Start the first run as a continuing process so local extraction and Figure work can continue while authoring proceeds:

   python scripts/run_pipeline.py --config pipeline-config.json --result pipeline-result.json
   

   Poll pipeline-progress.json. As soon as it reports early_evidence_ready, read early-evidence-packet.json and author every non-result slide into non-result-deck-spec.json. Do this concurrently with pdffigures2, Figure recovery, panel-coordinate handling, and Figure splitting. Do not author any result or results slide from the provisional packet, and do not change any content, length, layout, or speaker-note rule elsewhere in this skill.

3. Derive the complete Figure inventory from all main Figure legends and their text mentions, preserving paper source order. Run pdffigures2 once and assess whether it completely extracted every listed Figure. For each missing or rejected Figure, use the #C8EDCC alpha-mask recovery defined below and crop from the untouched original 300 DPI render. Recover independent Figures with at most four local workers. Reuse a rendered caption or adjacent page within the same run, reuse a verified Figure recovery across runs when the PDF, legend, and recovery-code fingerprints match, and leave large contrast diagnostics disabled unless a recovery failure is being investigated. If recovery still fails, request all unresolved Figures together. Treat every user-provided Figure as trusted complete: copy it unchanged, skip all completeness checks, and preserve PNG alpha.

4. Parse each Figure legend before panel processing. A Figure with fewer than four labels remains whole and performs no AI-vision or splitting work. After every Figure has an image, treat panel_label_vision_required as one internal model handoff. Read panel-label-vision-request.json; submit its contact sheet first and every listed dense Figure as an independent original-resolution input in source order; then write one schema-version-3 response to response_path. Each record contains only figure_id and label_boxes for the expected case-sensitive A-Z/a-z labels. Require each box to tightly enclose only the visible glyph; top is its topmost visible pixel, not its center, baseline, or a nearby panel-content boundary. Return all coordinates directly from that one pass; do not create per-label crops, perform glyph-by-glyph confirmation, or make a second vision request. Rerun the same work directory. Python does not classify Figure layout. It groups normalized label top anchors into horizontal rows whose total span is at most 0.2% of Figure height. Every row after the first, including a single-label row, anchors one line 0.3% of Figure height above its highest label. Python keeps candidates with at least 92% near-white pixels. Before whole-Figure fallback, use OpenCV to correct only a failed candidate by searching from 6% of Figure height above through 1% below for a continuous full-width run that is at least 99.99% near-white; center the candidate in the nearest qualifying run and apply the 92% gate again. Never refine an already-safe candidate or perform OCR. Select the surviving candidate closest to 50% of image height and directly write the crops only when both proposed panel groups remain contiguous in legend order; otherwise preserve the whole Figure so result ranges stay exact. Process independent Figures with at most four local threads after the batch coordinates arrive, then aggregate them in paper source order. If none passes, preserve the complete Figure. Python never recognizes label characters and never calls an AI API. During red-line testing, generate all final previews in one local batch and inspect only those previews. Intermediate per-label diagnostics are allowed only after a final line is demonstrably wrong.

5. Stop on failed. On user_intervention_required, read the single interaction-requests.json. It contains workflow_request and figure_requests; either may be null or empty. Ask the user once for the original-paper workflow decision and all unresolved Figure crops together. After the batch is returned, set workflow_decision, add workflow_image_path only for present, add Figure paths under reviewed_figure_overrides, and rerun the same work directory. Do not retry pdffigures2 when pipeline-state.json has the same input-PDF SHA-256. The single batch must include all unresolved Figures; ask the user to return all requested crops together.

6. Treat deck_spec_required as the result-slide handoff. Confirm that the final evidence_packet.json exists, author only the result slides that were held back, then merge them with the already authored non-result spec by running scripts/merge_deck_spec_parts.py. Renumbering is mechanical; scientific text, ordering rules, and speaker-note requirements must remain unchanged.

7. Read references/narrative-rules.md and choose the original-research or review-paper framework. Around 20 slides is a target, not a fixed limit.

8. Read references/deck-contract.md and produce deck_spec.json as the model's only authoring output. Every substantive slide except the cover must use exact anchors present in the evidence packet.

9. Add deck_spec_path, Qi PPTX, style-token path or optional style PPTX, output PPTX, and local Node path to the config; then rerun the pipeline. The pipeline must call validate_deck_spec and write deck-validation.json before compilation. Stop on deck_validation_failed; never invoke the compiler for an invalid deck. Use scripts/validate_deck_spec.py only for an optional standalone precheck; the in-pipeline gate remains mandatory. The precompile gate must also reject predicted overflow in fixed no-autoshrink boxes before consuming a compile attempt.

10. The pipeline writes figure-manifest.json from all main Figure render_assets. The compiler consumes those exact paths and orders; it does not scan directories or guess from filenames. Preload independent slide image assets in batches of at most four and export slide PNG/layout artifacts in batches of at most four. Artifact-tool slide mutations share one presentation and must not be moved into separate template copies when that would add merge or transfer overhead. Reuse unchanged PNG/layout artifacts by slide fingerprint, and reuse the entire compiled deck only when the deck spec, template, style tokens, Figure manifest and assets, paper header, and compiler fingerprint all match a prior QA-valid output.

11. Deliver only when pipeline-result.json reports complete and qa-report.json is valid. Preserve timing.json, render PNGs, and layout JSON for verification.

Runtime boundary

• Keep production conversion and skill maintenance separate. Freeze the skill during a full PDF-to-PPT run; do not patch code and silently restart the full workflow when a defect is discovered.
• Test a splitting-rule change only on the affected Figures in D:\codextest. Do not re-extract the PDF or compile the deck until the user explicitly asks for a new full run.
• Allow no more than 5 minutes without a new pipeline state. Stop and report the current state, elapsed time, and diagnostic path instead of continuing hidden debugging.
• Use 15 minutes as the default end-to-end wall-clock budget for an unchanged, dependency-ready local paper. If it is exceeded, stop before any new retry, report the slow stage from timing.json, and request permission before expanding into skill debugging. User-response waiting time is excluded.
• Reuse fingerprint-valid extraction, recovery, coordinate, and compile caches. Never invalidate an unaffected stage merely to retest a local rule.

Compilation budget

• 正式PPT只允许通过 scripts/run_pipeline.py 编译；禁止直接调用 compile_deck.mjs 绕过编译预算或质量门禁。
• 每个工作目录只允许首次编译和最多一次集中修正版编译；禁止第三次编译。
• 首次编译后，检查全部渲染页后再修改任何内容。把所有页面的检查结果和需要 一次性完成的修改写入 visual-review.json。只有该文件覆盖全部幻灯片、明确 标记已准备修订，而且 deck spec 或编译器已经实际变化时，流水线才允许第二次 编译。
• 如果首次输出通过检查，直接交付，不执行第二次编译。如果集中修正版仍不合格， 停止并报告剩余问题，不得继续循环生成。
• 最终QA不得触发额外编译；它只核验当前最终产物。

Figure rules

Use every existing main Figure discovered from all main Figure legends; there is no manual Figure-selection handoff. Run pdffigures2 once, verify each corresponding whole-Figure extraction, and send only missing or rejected Figures through background-contrast recovery. Render the caption page, then the previous and next pages with alpha; composite transparency onto diagnostic #C8EDCC for Python detection; crop accepted candidates from the untouched original 300 DPI render with a small label safety margin. Request all unresolved Figures together. A user-provided Figure is trusted complete, skips all completeness checks, and must preserve PNG alpha.

Read references/figure-splitting.md completely before panel processing. Derive the case-sensitive A-Z/a-z label set from each legend. Keep a Figure whole when the legend declares fewer than four labels. Put every remaining Figure into one batch_panel_label_coordinates request after all images are present. Supply one contact sheet and each dense Figure as an independent original-resolution input in paper source order. The schema-version-3 response contains one figure_id and its label_boxes; AI supplies coordinates only. Every box tightly encloses the visible label glyph, and top is the topmost visible glyph pixel rather than its center, baseline, or a nearby panel-content boundary.

Python does not classify Figure layout and generates horizontal candidates only. Group labels by normalized top; one row may span at most 0.2% of Figure height. For each row after the first, including a single-label row, place a candidate 0.3% of Figure height above that row's highest label. Keep candidates that pass the 92% gate and locally correct only a failed coordinate with the strict OpenCV white-run rule in references/figure-splitting.md, then select the survivor closest to 50% of image height. Directly write the upper and lower crops; if no candidate passes, preserve the whole Figure with no_safe_horizontal_split. For a trusted user-provided Figure, also preserve the whole image if coordinate validation or writing fails; do not request that Figure again. Preserve axes, legends, annotations, colorbars, scale bars, data, source channel count, and PNG alpha.

Evidence and web boundary

Keep all paper claims, conclusions, Figure explanations, and speaker notes grounded in the local PDF evidence packet. Web lookup is not evidence for paper claims. Use web lookup only for bibliographic and journal metadata: verify the latest JCR Journal Impact Factor from an official current journal or publisher metrics page, and record its source URL and lookup date; missing title, authors, journal, year, DOI, or publication date may be verified from official records. Record metadata sources in bibliographic-sources.json.

Never browse for or automatically populate research direction or author representative works. Render 待填写 in both named areas so the user can fill them after independent verification.

Speaker notes and delivery

Every slide requires complete Chinese prose that can be read aloud verbatim. Notes must explain the slide claim, evidence, Figure interpretation when present, and the transition to the next slide. Do not use short reminders such as “介绍研究背景”“讲一下结果” or “待补充”.

Write speaker notes as a spoken narrative rather than an expanded copy of the visible bullets. Keep the deck-wide average at or above 120 Chinese characters per slide, and meet the role-specific minimums in references/deck-contract.md. Preserve English paper titles, gene names, methods, pathways, and statistical terms where they improve precision, but explain their role in Chinese. Ground every scientific statement in the local evidence packet; never fill length with unsupported background.

• For the cover, identify the journal and exact paper title, state the paper's central evidence chain and proposed model, then preview two or three questions that organize the presentation.
• For the background, connect the clinical or scientific challenge to the opportunity, the unresolved gap, and the paper's core question. Do not merely read the four visible cards.
• For research_idea, narrate how the authors move from discovery to validation, localization, and mechanistic interpretation, using the paper's actual study order rather than generic stage names.
• For technical_route, explain all four stages separately. For each stage, state what the authors did, which evidence it produced, and why that evidence was needed for the next stage.
• For every result slide, first state the analytical purpose of the displayed Figure asset. Then name and explain every displayed panel separately, even when the sidebar merges adjacent panels. For each panel, briefly cover what is shown or encoded, the comparison or readout, and the evidence-supported finding. End with a distinct synthesis paragraph beginning with language such as 总体来看 or 综合来看, then transition to the next result. Never replace individual panel explanations with only a range such as Fig.2A-D.
• For conclusions, innovations, limitations, review synthesis, and outlook slides, explain the claim, its strongest evidence, its interpretation or boundary, and its relevance to the paper's overall question before the transition.

Compilation must duplicate mapped Qi frames and edit inherited text and image targets in place. Do not construct a parallel layout over the copied frame. The compiler does not author scientific content: every scientific title and body in structured layouts must come from validated content_blocks. Final QA must verify slide count, notes, hidden-but-preserved source metadata, the absence of visible source boxes, Figure assets, allowed blank content, structural placeholders, render resolution, layout bounds, and embedded image resolution. Use scripts/qa_deck.py directly for a standalone QA run or let scripts/run_pipeline.py invoke the same checks.