cyber-ppt

CyberPPT

Overview

Transform source materials into an evidence-chained, conclusion-first, editable consulting-style presentation. Review evidence, argumentation, and page architecture using MBB-level standards. Preserve evidence traceability, set confirmation gates, and judge completion by rendered results rather than file generation.

Mandatory Process

Stage	Required Output	Stop Condition	Read
1. Analysis	MBB-standard evidence table, conflict log, content brainstorming substep, SCR argumentation, page-by-page outline, chart plan, page information density, and component checklist	First confirmation: the user approves the storyline, page count, outline, per-page information structure, and density targets	references/source-analysis.md, references/storyline.md
2. Blueprint	8 visual styles, selected style, and a page-by-page ImageGen blueprint covering all requested/planned pages	Second confirmation: the user approves the visual direction and all page blueprints	references/visual-system.md
3. Hybrid Restoration PPTX	Adopt a hybrid restoration strategy of "complex visual fidelity + editable primary text," output PPTX, full-page rendered images, comparison checks, and editability notes	Final confirmation: the user approves the editable PPT	references/ppt-production.md, references/quality-assurance.md

Do not cross a confirmation gate without confirmation. When the user requests changes, return to the corresponding stage to revise and reconfirm.

Reference Gate

Before each stage begins, you must read all reference files in the "Read" column of the table above, and convert the key constraints of that stage into an execution checklist before acting. Do not execute based only on the main file summary, memory, existing scripts, or previous-round experience.

• Before a stage begins, you must read the complete content of the corresponding reference; if the terminal displays garbled text, reread it in UTF-8 instead, and do not skip it.
• Specific checklists in the reference take precedence over the summary descriptions in this file; if the two conflict, stop first, explain the conflict, and request user confirmation.
• In the first stage, you must read source-analysis.md and storyline.md before producing the evidence table, content brainstorm, SCR, page-by-page outline, and page information density checklist.
• In the second stage, you must read visual-system.md before generating style samples; by default, you must use the fixed 8 CyberPPT visual styles item by item, and must not substitute extended styles unless the user explicitly requests a replacement.
• In the page-by-page blueprint substage of the second stage, even if the style has already been selected, you must again compare it against visual-system.md, and declare the locked style number, color palette, grid, title hierarchy, chart language, and information density rules to prevent style drift during page-by-page generation.
• In the third stage, you must read ppt-production.md and quality-assurance.md before generating the PPTX and rendering checks.

Step 1: Evidence Analysis + Content Brainstorm + SCR + Page Density Planning

Step 1 is not "give a single-version outline after reading the materials." It must first turn the source materials into an auditable evidence base table, then use brainstorming to diverge into multiple optional storylines, and finally converge into the SCR, page-by-page outline, chart plan, and page information density rules.

Hard Requirements for Step 1

1. All facts, numbers, judgments, recommendations, caveats, and SO WHATs must be traceable to a location in the source materials or clearly marked as gaps.
2. An MBB standard evidence table must be established, recording source location, period, unit, confidence, conflicts, caveats, implications, and recommended visuals.
3. The content brainstorming sub-step must be completed: propose 2-3 optional storylines, build an issue tree or hypothesis tree, and compare evidence strength, risk gaps, applicable audiences, and reasons for non-recommendation.
4. A page material pool must be formed: key numbers, comparisons, rankings, changes, funnels, matrices, tables, annotations, legends, micro-charts, sidebar insights, and usable evidence IDs.
5. A recommended storyline must be output, rather than only providing one seemingly reasonable outline.
6. Information density and a component list must be defined for each page, including the number of information zones, the ratio of the main visual to the sidebar, the number of tables/annotations/legends/micro-charts, evidence IDs, SO WHAT, and low-density risk.
7. The user must not be asked only to confirm page titles; the first confirmation must cover the storyline, page count, page-by-page arguments, chart plan, information density, and component list.

Step 1 Work Sequence

1. Inventory input files and material roles.
2. Extract facts, numbers, tables, graphics, arguments, dates, units, and sources.
3. Establish an evidence table, and mark conflicts, gaps, and directional-only judgments.
4. Conduct content brainstorming based on the evidence table, proposing 2-3 optional storylines and an issue tree / hypothesis tree.
5. Compare storylines, choose the recommended path, and explain the trade-offs.
6. Converge the recommended path into a full-document SCR.
7. Generate a page-by-page plan: conclusion title, role, detailed argumentation, evidence, caveats, chart plan, visuals, implications, and transitions.
8. Complete the page information density and component list for each page as input for the Step 2 blueprint.

Step 1 Confirmation Output

The first confirmation must include:

• MBB evidence table summary;
• open data conflicts, missing evidence, and caveats;
• 2-3 optional storylines and the reasons for recommendation;
• recommended SCR;
• page-by-page outline and chart plan;
• page-by-page information density and component list;
• data questions requiring user decisions or supplements.

Step 1 Prohibitions

• Do not skip content brainstorming and directly submit a single-version outline.
• Do not use common sense to fill in data that the source materials do not contain.
• Do not treat ImageGen text, numbers, logos, quotations, or chart values as evidence.
• Do not output a low-density outline containing only 'page title + brief description'.
• Do not leave low-density risks for the blueprint stage to resolve on its own.

Step 2: 8 Fixed Sample Templates + Page-by-Page ImageGen Blueprint

Step 2 is not a text style choice, nor is it merely writing a style specification. It is divided into two sub-stages: first, actually display 8 fixed CyberPPT visual sample templates for the user to choose from, then generate a complete page-by-page ImageGen blueprint for the full page count according to the selected style and the page density confirmed in Step 1.

Step 2 Hard Requirements

1. Must read visual-system.md。
2. Must directly send 8 independent and complete 16:9 sample images through the current conversation for the user to choose from. Prefer using built-in sample templates assets/palette-samples/palette-01.png to palette-08.png; if regenerating, 8 real images must also be delivered.
3. Web pages, HTML, URLs, folder paths, file lists, Markdown tables, text descriptions, collages, or thumbnail walls can only serve as supplements and cannot replace the 8 independent sample images in the current conversation.
4. If web assistance is used, the web page may only serve as an additional browsing method and must not be the sole basis for style confirmation.
5. Before the user can directly see the 8 sample images in the current conversation, you must not ask the user to choose a style, and you must not enter the page-by-page blueprint stage.
6. You must not generate only Markdown, tables, text lists, recommendation rationales, or stage2_style_options.md as the style confirmation artifact.
7. By default, the fixed 8 CyberPPT visual styles must be used; unless the user explicitly requests a replacement, they must not be changed to improvised extended styles.
8. After a style is selected, the style number, name, color palette, grid, heading hierarchy, chart language, header and footer, and information density rules must be declared and locked.
9. The page-by-page blueprint must use the page information density and component list confirmed in the first step, and must not downgrade a high-density plan into loose cards, generalized charts, or sparse large-text posters.
10. The text and numbers in the blueprint images still must not serve as the final source of truth; however, before the page-by-page blueprint is generated, slide_content_lockmust first be established to lock the page content using the real text, data, table structures, SO WHAT, evidence IDs, and required components confirmed in the first step. The third stage must fill in the real content from this lock file, and must not reinterpret the source material, remove components, or reduce information density.
11. The second stage must automatically determine the default target delivery language, and must not add a separate confirmation step for language selection. The priority for default language determination is: the global delivery language explicitly specified by the user > the primary language of the source material > the language of the current conversation. Ask the user only when the source material is multilingual with no obvious primary language, or when the user's instructions conflict with the language of the source material.
12. The page-by-page ImageGen blueprint must use the effective target language. Effective target language = default target delivery language + registered page-level, section-level, or component-level language overrides. Titles, module titles, chart labels, legends, axis labels, annotations, sources, footers, SO WHAT, and other visible text placeholders in the blueprint must all use the effective target language of the corresponding scope; English blueprints must not be generated by default.
13. If the user explicitly requests that a certain page, section, or component use a different language, it must be recorded as language_overrides, and the override language must be used only within the corresponding scope. The entire blueprint must not be changed to English because of a local English requirement.
14. target_language、language_source、effective_language、language_overrides and allowed_foreign_terms are execution metadata and may only be written in blueprint records, prompt descriptions, manifests, or QA records, and must not appear as visible page text in the blueprint images or the final PPT.
15. After each page's blueprint is confirmed, blueprint_component_signaturemust be frozen, recording the component type, component structure, subcomponents, and corresponding content lock references. The third stage may only read it, and must not rewrite, relax, or temporarily add signatures.
16. After each page's blueprint is confirmed, visual_element_registrymust be established, registering the bbox, type, priority, owning component, and tolerance of all visible elements in the blueprint. P0/P1/P2 only determine tolerance and failure level, not whether measurement can be skipped.
17. slide_content_lock、blueprint_component_signature and visual_element_registry must record the file path and SHA-256; the third-stage manifest may only reference the frozen results and must not rewrite them on its own.

Second-Step Work Order

1. Send the 8 fixed sample images directly through the current conversation, and list the number, name, color palette, tone, strengths, and risks outside the images.
2. Wait for the user to choose or request a style replacement.
3. Lock the visual system: page size, safe margins, grid, font hierarchy, chart language, table style, header and footer, accent colors, and density rules.
4. Based on the page-by-page plan confirmed in step one, generate one complete 16:9 ImageGen blueprint for each page; unless the user explicitly skips ImageGen, script drawing, PPT, HTML, SVG, canvas, or low-fidelity mockups must not be used as substitutes.
5. Before each page blueprint, first generate slide_content_lock; after each page blueprint is confirmed, freeze blueprint_component_signature and visual_element_registry, and record the path, SHA-256, page role, complex visual asset areas, editable text areas, native component list, evidence ID, and expected information density.
6. Check for style drift, decreased information density, overly dense covers, overly sparse content pages, and generated text contamination.

Step Two Confirmation Output

The second confirmation must include:

• 8 independent 16:9 style sample images;
• Style number, name, color palette, applicable tone, advantages, and risks;
• A visual system lock-in statement after the user makes a selection;
• Page-by-page ImageGen blueprints covering all pages;
• Each page blueprint's ImageGen prompt, output image path or generation record,slide_content_lock、blueprint_component_signature、visual_element_registry, component list, editable text areas, complex visual asset areas, evidence ID, and density target;
• Style drift or information density risks that require user confirmation.

Step Two Prohibitions

• Must not provide only stage2_style_options.md or a text description to let the user choose a style.
• Must not use a collage, thumbnail wall, or compressed multi-page canvas to replace the 8 independent samples.
• Must not provide only a webpage, HTML, URL, folder path, or file list for the user to view the samples; these may only serve as a supplement after images are displayed in the current conversation.
• Must not ask the user to choose a style when images on the webpage have not loaded, paths are invalid, or the user cannot see the samples.
• Must not enter the page-by-page blueprint stage before the user has confirmed that they can see the 8 samples.
• Must not interpret "only use source files" as skipping visual samples or skipping ImageGen; it only constrains the source of facts, data, and text.
• Must not let ImageGen add content density on its own; content density must come from the page plan confirmed in step one.
• Must not diverge again into an unselected visual system during the page-by-page blueprint stage.
• Must not use PptxGenJS, python-pptx, HTML, CSS, SVG, canvas, Pillow, matplotlib, PowerPoint, or any local drawing script to directly draw page-by-page blueprints. Except for python-pptx , these tools may only be used for prompt management, metadata, QA, reference images, non-generation assistance in stage three, or assistance workflows after the user explicitly skips ImageGen;python-pptx must not be used for formal PPTX generation in stage three.
• Must not pass off PowerPoint pages, webpage screenshots, wireframes, structural sketches, default card pages, low-fidelity mockups, or regular placeholder images designed for easy measurement as ImageGen blueprints.
• If the user explicitly skips ImageGen, must record imagegen_skipped_by_user=true, the path to the user-provided template/screenshot/brand guidelines/visual specification, and the basis for substitution; must not claim that the page is an ImageGen blueprint.
• visual_element_inventory_targets and blueprint_measurement_targets are only stage-two record metadata, used for stage-three restoration preparation; blueprints must not therefore be downgraded into script sketches, wireframes, default card pages, or low-fidelity dashboards.

Step Three: Complex Visual Fidelity + Main Text Editable

Adopt a hybrid restoration strategy of "complex visual fidelity + editable primary text". The highest priority is to preserve the overall design texture, composition, visual completeness, and premium feel of the original image, while ensuring that the main text content can be edited directly in PPT. Do not simply lay out the entire page as a single background image.

Hard requirements

1. Primary text must not be retained as a whole as an image; it must be rebuilt using native PPT text boxes.
2. Text that must be editable includes: main title, subtitle, body paragraphs, key quotes, key numbers, keywords, conclusion box text, annotation text, footer notes, section names, and page numbers.
3. Small text inside complex visual areas may be retained as an image if separating it would obviously damage the visual texture, but it must be noted in the final description.
4. Do not compress all text into images for the sake of visual fidelity.
5. Do not rebuild complex visual effects into low-quality default PPT shapes for the sake of full editability.
6. The ImageGen blueprint may only be used as a reference for composition, hierarchy, density, and visual language; the full-page blueprint or large-area blueprint screenshots must not be used as the final PPTX background.
7. Line charts, bar charts, axes, labels, key numbers, tables, comparison bars, process arrows, SO WHAT, and headers and footers in content pages must by default be rebuilt using native PowerPoint text and shapes.
8. If a page is confirmed after complex visual scanning to have no complex visual assets, and the blueprint allows fully native reconstruction, then that page can usually ultimately achieve pictures=0; otherwise, the necessity of images, SVG, custom geometry, or freeform assets must be explained item by item.
9. Phase 3 delivery is not "generate PPTX + strict QA", but "page-by-page blueprint comparison passed + PPTX structure passed". If either fails, delivery confirmation must not be given.
10. The approved blueprint is the visual acceptance baseline, not an inspiration image or a content structure reference. Page background color, surface system, header and footer, SO WHAT, column relationships, main image form, icon anchors, density, and whitespace rhythm must all be compared item by item.
11. Before generating the PPTX for each page, you must first write blueprint_reconstruction_plan, breaking down the blueprint's layout, density, surface system, and anchors. Without this record, the page must not be generated.

Phase 3 user confirmation prompt

When entering Phase 3 and waiting for user confirmation, you must state in a prominent position:

Next is not generating the complete PPT all at once, but producing and accepting page by page. This is to prevent the AI from losing focus during batch generation, which would lead to decreased information density, drift in curve/icon/text positions, inconsistency in the background surface system, and loss of visual semantics. Only after each page is confirmed and passed do you proceed to the next page; after all pages pass, merge the approved single-page PPTX files into the complete deck, and perform post-merge regression verification.

Phase 3 page-by-page execution and page-by-page acceptance gate

Phase 3 must by default be executed page by page; the complete deck must not be generated all at once as the final version.

Each page must be executed in order:

1. Read only the current page's blueprint;
2. Output the current page's blueprint_reconstruction_plan；
3. Complete the current page's complex_visual_scan；
4. Determine the admission of images, SVG, custom geometry, and complex visual assets;
5. Determine whether the precise curve, pixel boundary tracing, table density, position anchor, font size, and overflow gates are triggered;
6. Generate only the current page's single-page PPTX;
7. Export the current page's rendered image;
8. Generate the current page's blueprint vs render side-by-side comparison image;
9. Perform local comparisons on key areas such as the title area, main chart area, process/icon area, SO WHAT, and footer;
10. Fill in the current page manifest and visual QA evidence;
11. Fix the current page until it passes;
12. Only after the user confirms the current page may you proceed to the next page.

Before the current page passes, you must not produce the next page. When the user requests "high fidelity," "1:1," "restore according to the blueprint," "formal delivery," or "precise restoration," Stage 3 prohibits generating 2 or more pages at once as the final version. Multiple pages generated at once may only be marked as a rough draft and must not be marked as a CyberPPT qualified final version. You must not first batch-generate the complete PPTX and then afterward backfill the manifest, visual QA, or side-by-side to disguise it as page-by-page acceptance; visual QA must be generated immediately based on the actual rendering results of the current page.

Final merge gate

The final complete PPTX must be obtained by merging single-page PPTXs that have already passed page-by-page acceptance.

During the merge stage, you must not rerun the page generation scripts to rebuild pages, must not re-layout, must not redraw charts, must not reapply backgrounds, must not use single-page rendered images as full-page backgrounds, and must not replace pages that have passed with native objects using screenshots.

During the merge stage, you may only import single-page PPTX pages that have passed acceptance, and you should preserve the original page XML, theme, background, shapes, text, charts, SVG/custom geometry, and relationship files as much as possible. After merging, you must export PNGs for regression verification; rendering here is only QA, not re-making the pages.

Each page after merging must be compared against the corresponding accepted single-page rendering. If there is a background color change, page size change, object offset, font change, missing image, SVG distortion, chart misalignment, or decrease in information density, the merge has failed, and you must fix the merge process rather than regenerate the pages.

Dual hard thresholds: editability and visual semantics must both hold

结构可编辑 and 视觉还原 are equally hard thresholds, not an either-or choice, and must not override each other. The final PPTX must simultaneously satisfy "information structure is editable" and "visual semantics are high-fidelity."

Hard criteria:

• You must not use editability as a reason to simplify, replace, or downgrade key visual expressions in the blueprint. For example, you must not change flow lines, arcs, irregular shapes, spatial relationships, unified surface systems, color hierarchies, or visual focal points into ordinary rectangles, default flowcharts, stacked white cards, or rough lines.
• You must not use visual fidelity as a reason to turn main text, key numbers, chart labels, SO WHAT, headers and footers, or simple charts into images.
• When editability and visual fidelity conflict, you must adopt a hybrid approach of "native objects + small-area text-free SVG/custom geometry/tightly cropped non-text assets + manifest disclosure," rather than sacrificing either goal.
• If pictures=0 but the key chart form, curve amplitude, panel logic, background color system, or visual semantics clearly do not conform to the blueprint, it still fails.
• If images/SVGs are used to carry main text, numbers, labels, sources, footers, or SO WHAT, it still fails.
• qa_expectations must declare dual_gate_required: true、visual_semantics_required: true and all_key_text_editable: true; if missing, it is considered a Stage 3 failure.
• visual_qa_gate.json must separately provide editable_information_layer_pass and visual_semantics_preserved. If either is false，deliverable_allowed must be false。

pictures=0 non-goal principle

pictures=0 is not a Stage 3 goal, nor is it proof of visual qualification.

Stage 3 has only two equally hard thresholds:

1. The main information layer is editable;
2. blueprint visual semantics are high-fidelity.

For pages without complex visual assets, PowerPoint native objects should be used preferentially for reconstruction, and the final result can usually reach pictures=0。

However, when photos, official assets, complex textures, complex illustrations, complex icons, flow lines, irregular boundaries, complex arcs, non-standard chart forms, or non-text visual assets in the blueprint constitute visual semantics, the asset admission gate or precise tracing gate must be triggered, and a solution must be chosen that preserves both visual semantics and the editability of the main information. Small text-free images, SVG paths, custom geometry, or high-density freeform may be used; images/SVG must not carry the main text, key numbers, chart labels, SO WHAT, headers and footers, or sources.

It is not permitted, for the sake of pictures=0：

• changing a flow diagram into rectangular bars;
• changing an arc process into an ordinary polyline;
• changing complex icons into default icons;
• changing a unified page surface system into white cards;
• reducing page density;
• simplifying the blueprint's visual focus;
• actively avoiding triggering the image, curve, irregular-shape, or complex visual gates.

Restoration strategy

1. First identify the page structure, and split the composition into a "complex visual asset layer" and an "editable information layer".
2. The complex visual asset layer should preferentially be cropped/extracted from the original image as high-resolution image assets and embedded into the PPTX.
3. The editable information layer should be reconstructed using PPT native text, shapes, lines, basic graphics, and simple charts.
4. If the original image text needs to be changed into editable text, please use background color, gradient approximation, or local masking to cover the original image text, then overlay editable text to avoid ghosting.
5. For complex icons, brand logos, photos, illustrations, 3D images, collages, complex charts, complex flowcharts, complex cycle diagrams, complex backgrounds, materials, textures, lighting and shadow, depth of field, glassmorphism, metallic texture, soft light, and shadows, they may be retained as images.
6. For simple lines, dividers, basic geometric frames, lightweight labels, simple buttons, simple color blocks, page numbers, basic tables, and simple charts, they must be reconstructed as editable PPT elements.
7. Brand logos, trademarks, product UIs, people, product images, and official icons must not be hand-drawn or faked; they should preferentially be retained as cropped image assets from the original image.
8. If the chart data cannot be accurately inferred from the image, please reconstruct it approximately according to visual proportions, and note "data is a visual approximation" in the description.

PPTX generation tool selection gate

The formal PPTX pages in Phase 3 must be generated using PptxGenJS / pptx-generator. PptxGenJS is the carrying and layout engine, used to place native text, rectangles, tables, basic lines, chart components, traced SVG paths, PPT custom geometry, and high-density freeform; it is not permission to downgrade complex visuals into PowerPoint preset shapes.

In the formal PPTX generation of Phase 3, the use of python-pptxis strictly prohibited. Regardless of whether PptxGenJS reports errors, whether PowerPoint can open it, whether the task is urgent, or whether the page is simple, it is not permitted to switch to python-pptx, HTML-to-PPT, screenshot-to-PPT, or other generation engines.

Each page manifest must record generation_engine.tool="pptxgenjs" or the equivalent pptx-generator, and declare visual_fidelity_not_reduced=true。generation_engine.tool as python-pptx、python_pptx、html-to-ppt、screenshot-to-ppt or other formal generation engines, all of which are regarded as Phase 3 failures.

SVG / custom geometry priority gate

When the blueprint contains curves, arcs, sectors, annular gaps, flow bands, irregular boundaries, non-rectangular regions, complex icon outlines, or user-required 1:1 geometry-sensitive visuals, SVG path, PPT custom geometry, or high-density freeform must be used preferentially for reproduction.

Icon Library Priority Gate

When the third stage involves generic icons, badges, in-node symbols, SO WHAT icons, or explanation bar icons, verified SVG icons must be preferentially selected from assets/icons/ , and complex SVG icons must not be hand-written on the fly. The icon library is used to avoid broken lines, discontinuities, rough paths, and style drift; icons are not required to be exactly identical to the random icons in the ImageGen blueprint, but they must be semantically approximate, stylistically unified, and consistent in line width/fill, and must be entered into visual_element_registry for coordinate reverse testing.

Icon library usage rules:

• Only one stylistic library may be selected for the generic icons of each PPT set:chunk-filled、tabler-filled、tabler-outline or phosphor-duotone. Multiple generic icon styles must not be mixed.
• simple-icons It is only allowed for real brand, company, product, or service logos, and must not be used as a substitute for ordinary concept icons.
• Before selecting an icon, scripts/select_icon.py search --index assets/icons/index.json --query "<keyword>" --library "<library>" must be used to search candidates; do not read the entire icon library.
• After selecting an icon, the manifest or production record must register icon_id、source_library、semantic_meaning、svg_path、viewBox and powerpoint_render_verified。
• Icons must still register blueprint_bbox_px、ppt_target_bbox_in、render_bbox_px、delta_px and tolerance_px; spatial registration must not be skipped because an icon library is used.

Only when there is no semantically approximate icon in the icon library is custom_icon_required=trueallowed, and the SVG path / custom geometry precise tracing gate must be followed. Complex icon SVGs that are broken, rough, or unrenderable must not be hand-written on the fly just to "look more like the blueprint."

PowerPoint preset shapes must not be used to replace precise tracing results, including but not limited to:

• pie；
• arc；
• blockArc；
• chord；
• moon；
• wave；
• complex arrows;
• default flowchart shapes;
• connector line segment stitching.

SVG path, PPT custom geometry, and high-density freeform are the preferred implementation methods for geometry-sensitive visuals.pictures=0, editability, PptxGenJS native objects, or PowerPoint preset shapes cannot override this gate.

Preset Shape Admission Gate

Complex PowerPoint preset shapes must not enter formal pages by default. They are allowed only when all of the following conditions are met:

1. Single-object PowerPoint open and PNG export testing has been completed;
2. A local blueprint vs render comparison has been completed;
3. The visual error is below the tolerance for that region;
4. The manifest records why SVG path/custom geometry is not used;
5. The object will not cause zero size, negative size, negative coordinates, illegal adjustment parameters, or PowerPoint being unable to open.

Otherwise, SVG path, custom geometry, high-density freeform, or a stable combination of native objects must be used instead.

PowerPoint Compatibility and Corruption Handling Gate

After each page of PPTX is generated, it must be opened with PowerPoint and exported as PNG. Only when the complete-page PPTX can be opened by PowerPoint and the current page PNG can be successfully exported is it allowed to enter visual QA, manifest approved status, and user confirmation. Passing ZIP/structure pre-checks does not equal passing PowerPoint compatibility.

If PowerPoint cannot open the page, the following must be executed in order:

1. Delete the current page's old PPTX, old PNG, and old QA files to prevent old files from remaining;
2. Scan the slide XML for zero size, negative size, negative coordinates, abnormal ext/off, illegal transparency, and illegal line width;
3. Isolate objects group by group, and locate the object or object group that first causes PowerPoint to fail to open;
4. Replace the bad object with a PowerPoint-safe SVG path, custom geometry, high-density freeform, or a stable combination of native objects;
5. Regenerate the complete page;
6. Open it again with PowerPoint and export PNG.

Do not switch the generation engine before object isolation is complete; under no circumstances switch to python-pptx。

Isolation files are not deliverables

Blank PPTX, semi-finished PPTX, group test PPTX, and object isolation PPTX produced during the compatibility localization process may only be named isolation-* or compat-test-*, and may only be used to locate problems.

These files must not be named slide-XX.pptx, must not be written into the manifest, must not enter visual QA, must not be given to the user for confirmation, and must not be used as final pages or merge sources. Only after a complete blueprint page opens and renders in PowerPoint can it be written into pages/slide-XX.pptx。

Image asset admission gate

Before producing the PPTX for each page, an image asset admission table must first be established; without an admission table, generation of that page must not begin.

Area	Whether images are allowed by default	Determination
Title, subtitle, body text, key numbers, footer, page number, SO WHAT	No	Primary information layer, must be editable
Line charts, bar charts, axes, labels, comparison bars, simple process arrows, basic tables	No	Simple charts/basic graphics, must be rebuilt natively
Photos, official logos, product UI, complex illustrations, complex textures, 3D, glassmorphism, lighting and shadow materials	Allowed	Allowed only when native rebuilding would significantly degrade quality
Large-area or full-page blueprint screenshots	No	The blueprint is a reference, not a delivery background

Every image asset must record: source, reason for retention, coverage area, and whether editability is sacrificed. If necessity cannot be given, rebuild it with native objects instead.

slide_manifest.json hard threshold

Before each page enters PPTX generation in phase three, slide_manifest.jsonmust first be written, and after generation it must be updated according to the actual PPTX. No manifest, missing pages in the manifest, or incomplete manifest fields are all regarded as phase three failure, and delivery confirmation must not be made.

Each page manifest must contain at least:

{
  "slide": 2,
  "role": "Situation",
  "layout_reference": "blueprints/slide-02.png",
  "generation_engine": {
    "tool": "pptxgenjs",
    "fallback_reason": null,
    "visual_fidelity_not_reduced": true
  },
  "page_execution": {
    "mode": "single_page",
    "single_page_pptx_path": "pages/slide-02.pptx",
    "blueprint_render_path": "blueprints/slide-02.png",
    "ppt_render_path": "renders/slide-02.png",
    "side_by_side_path": "qa/slide-02-side-by-side.png",
    "local_comparison_artifacts": [
      "qa/slide-02-title-crop.png",
      "qa/slide-02-main-chart-crop.png",
      "qa/slide-02-so-what-crop.png"
    ],
    "page_status": "approved",
    "user_confirmed": true,
    "made_before_next_slide": true
  },
  "blueprint_reconstruction_plan": {
    "blueprint_path": "blueprints/slide-02.png",
    "canvas_size": "16:9",
    "background_color_sample": "#F3F4EF",
    "surface_system": "continuous paper with subtle panels",
    "layout_regions": ["title", "main_chart", "right_evidence", "so_what"],
    "header_footer_system": "page badge, source footer",
    "so_what_region": "bottom band",
    "main_chart_semantics": "line chart with evidence callouts",
    "density_targets": "high-density consulting page",
    "anchor_targets": ["title baseline", "chart origin", "right panel edges"],
    "native_rebuild_targets": ["title", "labels", "chart", "so_what"],
    "allowed_visual_assets": [],
    "complex_visual_scan": {
      "completed": true,
      "complex_visual_candidates": ["line chart callouts", "subtle panel system"],
      "triggered_gates": ["spatial_registration"],
      "native_only_rationale": null,
      "pictures_zero_is_not_goal": true
    }
  },
  "expected_pictures": 0,
  "image_assets": [],
  "text_objects": [
    {
      "id": "page_title",
      "role": "T2",
      "text": "页面主标题",
      "font_size_pt": 24,
      "editable": true
    }
  ],
  "native_components": [
    {
      "id": "main_chart",
      "type": "line-chart-shapes",
      "reason": "简单折线图，必须原生重建"
    }
  ],
  "qa_expectations": {
    "pictures_must_be_zero": true,
    "all_key_text_editable": true,
    "typography_scale_required": true,
    "dual_gate_required": true,
    "visual_semantics_required": true,
    "visual_qa_required": true,
    "spatial_registration_required": true,
    "container_overflow_check_required": true,
    "continuous_text_flow_check_required": true,
    "table_semantic_typography_required": true,
    "table_density_check_required": true
  }
}

Hard determination:

• Only after completing complex_visual_scan and confirming that there are no complex visual assets, no asset admission gate, and no precise tracking gate,expected_pictures may it be set to 0，image_assets may it be an empty array.
• As long as expected_pictures = 0 or pictures_must_be_zero = true, that page in the final PPTX pictures > 0 fails; but expected_pictures = 0 or pictures_must_be_zero = true must not be treated as a goal, and complex visual gates must not be actively avoided because of this.
• blueprint_reconstruction_plan.complex_visual_scan must record the scan completion status, complex visual candidates, triggered gates, native-only reasons, and pictures_zero_is_not_goal=true. Missing or incomplete means failure.
• generation_engine must record the PPTX generation tool and visual_fidelity_not_reduced=true; formal PPTX generation in phase three must use PptxGenJS / pptx-generator, and any python-pptx, HTML-to-PPT, screenshot-to-PPT, or other generation engine fails.
• page_execution The current page's single-page production, rendering, side-by-side, local comparison, user confirmation, and "only enter the next page after confirmation" status must be recorded. If missing or not mode=single_page , it fails.
• page_execution.page_status It must be approved，user_confirmed and made_before_next_slide must be true; otherwise, it must not proceed to the next page or final merge.
• If any image covers more than 40% of a single page, but image_assets does not explain the source, area, and necessity item by item, it fails.
• If any image covers more than 90% of a single page, it is handled as a full-page background risk; content pages fail by default, and "editable text is overlaid on top" may not be used as an exception.
• text_objects In role、font_size_pt and editablemust be stated for every major text object; if missing, it fails.
• role InC0, T1-T14); falling below the corresponding lower limit fails, and "for density" may not be used as an explanation.
• Line charts, bar charts, axes, labels, key numbers, tables, comparison bars, process arrows, SO WHAT, headers, and footers must be registered as native_components; if they are carried by images, it fails.
• qa_expectations.dual_gate_required、qa_expectations.visual_semantics_required and qa_expectations.all_key_text_editable must both be true; it is not allowed to declare only editability or only visual fidelity.
• For central diagrams, flowcharts, architecture diagrams, ecosystem diagrams, matrix diagrams, path diagrams, icon-dense pages, or pages containing combinations of nodes/arrows/icons/labels,qa_expectations.spatial_registration_required must be true, and spatial_registration_check。
• When any page contains cards, panels, tables, conclusion bars, SO WHAT, chart annotations, or fixed-area text,qa_expectations.container_overflow_check_required must be true, and container_overflow_check。
• When any semantically continuous sentence is split into multiple text boxes, multiple rich-text fragments, multiple highlighted segments, or cross-region text,qa_expectations.continuous_text_flow_check_required must be true, and continuous_text_flow_check。
• When any page contains tables, matrices, action lists, risk lists, or grid-based management tables,qa_expectations.table_semantic_typography_required and qa_expectations.table_density_check_required must be true, and table_text_objects and table_density_check。
• Violating the literal rules is violating the rules themselves; the manifest must not be bypassed on the grounds of "visual fidelity," "restore according to the blueprint," "the user only looks at the rendered image," or "it will be changed later."

Phase 3 High-Fidelity Blueprint Restoration Gate

Phase 3 must be the high-fidelity blueprint restoration phase; there is no "low-fidelity formal delivery," "ordinary restoration formal delivery," "delivery as long as it is semantically similar," or "one-off draft first."

Phase 3 must not first generate low-fidelity drafts, quick previews, rough drafts, or batch first drafts, and then retroactively fill in the manifest, visual QA, or user confirmation records.

All visible visual elements must enter visual_element_inventory, and be classified into item-by-item measurement, intra-group sub-anchor measurement, or decorative group measurement. Unregistered visible elements are regarded as omissions by default; before omissions are fixed, visual_semantics_preserved=truemust not be set, deliverable_allowed=truemust not be set, and delivery confirmation must not be given.

Registration must not be skipped because an element is "tiny," "decorative," or "does not affect text."

visual_element_inventory In priority：

Phase 3 must establish visual_element_registryfor all visible elements in the blueprint. P0/P1/P2 only affect tolerance and failure level; they do not affect whether registration, measurement, restoration, and post-render reverse measurement are required.

• P0: Title, main image, SO WHAT, footer, key numbers, core panels, user-indicated areas. Default tolerance 3px, maximum 6px; exceeding the tolerance is Critical.
• P1: Ordinary cards, icons, labels, arrows, tables, dividers. Default tolerance 4px, maximum 8px; exceeding the tolerance is High, and must not exist in formal delivery.
• P2: Decorative lines, dot patterns, textures, repeated ticks, background patterns. Default tolerance 6px, maximum 12px; exceeding the tolerance is Medium, and if it causes visible misalignment, ghosting, reduced density, or broken rhythm, it is upgraded to High/Critical.

All P0/P1/P2 elements must include element_id、priority、element_type、source_component_id、blueprint_bbox_px、ppt_target_bbox_in、render_bbox_px、delta_px、tolerance_px and registration_status. Registration or reverse testing must not be skipped on the grounds of "decoration", "P2", "unimportant", or "visually approximate".

Minimal example:

{
  "visual_element_inventory": [
    {
      "id": "title_block",
      "priority": "P0",
      "role": "title",
      "measurement_mode": "individual_bbox",
      "must_reproduce": true,
      "blueprint_bbox_px": {"x": 90, "y": 64, "w": 800, "h": 72},
      "ppt_target_bbox_in": {"x": 0.63, "y": 0.44, "w": 5.56, "h": 0.5},
      "tolerance_px": 3
    },
    {
      "id": "evidence_cards",
      "priority": "P1",
      "role": "cards",
      "measurement_mode": "group_with_child_anchors",
      "group_bbox_px": {"x": 1220, "y": 210, "w": 560, "h": 520},
      "child_anchors": [{"id": "card_01", "anchor": "top_left"}],
      "must_reproduce": true
    },
    {
      "id": "background_dot_pattern",
      "priority": "P2",
      "role": "micro_decoration",
      "measurement_mode": "decoration_group",
      "group_bbox_px": {"x": 80, "y": 140, "w": 1760, "h": 820},
      "color": "#D8DAD2",
      "density": "medium",
      "spacing_px": 12,
      "alignment": "grid",
      "repeat_direction": "xy",
      "opacity": 0.42,
      "reproduction_strategy": "native repeated dot pattern"
    }
  ]
}

Hard criteria:

• Missing visual_element_inventory means failure.
• Missing visual_element_registry means failure.
• Missing priority means failure;priority can only be P0、P1、P2。
• visual_element_registry Every P0/P1/P2 element in must provide a blueprint bbox, PPT target bbox, post-render bbox, delta, and tolerance.
• P0 Default tolerance 3px, maximum 6px;P1 default tolerance 4px, maximum 8px;P2 default tolerance 6px, maximum 12px.
• It is not allowed to register only containers, panels, the main chart area, or the SO WHAT large box to replace measurement of internal child elements.
• P0/P1 must not be downgraded to P2; marking the main image, title, SO WHAT, footer, key numbers, core panels, and user-indicated areas as P2 is considered failure.

Blueprint Measurement-Driven Restoration Gate

Before page-by-page generation in Phase 3, blueprint_measurement_tablemust first be established. This table is the execution basis for converting the approved blueprint into PPT coordinates, not explanatory text.

Canvas conversion fields must be recorded:

{
  "blueprint_canvas_px": {"w": 1920, "h": 1080},
  "ppt_canvas_in": {"w": 13.333, "h": 7.5},
  "scale_x": 0.006944,
  "scale_y": 0.006944
}

Each key area must record:

{
  "id": "main_chart_area",
  "priority": "P0",
  "role": "main_chart",
  "blueprint_bbox_px": {"x": 118, "y": 202, "w": 1320, "h": 610},
  "ppt_target_bbox_in": {"x": 0.82, "y": 1.40, "w": 9.17, "h": 4.24},
  "tolerance_px": 3,
  "must_match_geometry": true
}

Hard criteria:

• qa_expectations.visual_semantics_required=true When, missing blueprint_measurement_table means failure.
• Missing blueprint_canvas_px、ppt_canvas_in、scale_x or scale_y means failure.
• visual_element_inventory If not all visible elements are covered, it is failure.
• If P0 key semantic elements are not measured item by item, it is failure.
• If P1 elements are not measured item by item or within groups, it is failure.
• If P2 tiny repeated elements are not included in decorative group measurement, it is failure.
• Coordinate replication must not be replaced by "semantically similar", "structurally consistent", "roughly restored", or "layout reference".

Visual System Restoration Gate

Before generating the PPTX for each page in Phase 3, the visual system of the approved blueprint must first be converted into an executable record. Without the record, that page must not be generated.

Must record:

• Page background color, panel background color, chart area background color, conclusion bar background color, and footer background color;
• Whether the page is a continuous paper surface, whether independent white cards are allowed, and whether white is only local highlighting;
• Lines, borders, column headers, dividers, shadows, rounded corners, and whitespace rhythm;
• The visual semantics of the main chart: ordinary bar/line chart, table, matrix, process, migration, flow line, Sankey, arc, waveform, irregular region, or other;
• Whether the page triggers the curve/irregular-shape visual precise tracing gate.

Hard criteria:

• The blueprint adopts a unified page surface system: the entire page primarily relies on the selected style's background base color, panel color scale, thin borders, column headers, divider lines, whitespace, or slight light-dark differences to create zones. Therefore, the third stage must restore according to the page base color, panel base color, chart area base color, and zoning method recorded in the blueprint, and the final PPTX must not be arbitrarily rebuilt as a large-area stack of pure white cards.
• The panel base color must match the blueprint's visual system. Do not treat #FFFFFF as the default content area base color, unless the blueprint is explicitly white cards.
• validate_pptx.py --strict passing does not mean the visual system is qualified. When the base color logic, panel logic, chart semantics, or curve geometry in the rendered image clearly deviates from the blueprint, it is considered High/Critical, must be reworked, and must not be delivered for confirmation.
• Do not use "structure is editable," "QA passed," "no images," or "the user can modify it later" to explain visual system deviations.

Curve/Irregular Shape Visual Precise Tracing Gate

When curves, flow ribbons, irregular boundaries, non-standard arcs, or any geometry-sensitive visuals exist in the blueprint or reference image, the precise tracing gate must be triggered. After triggering, ordinary rectangles, parallelograms, default flowcharts, default wave shapes, or rough PPT curves must not be used for approximation.

Trigger scenarios include:

• Sankey diagrams, streamline diagrams, value migration diagrams, Ribbon, band flows, funnel flows, envelope area bands;
• Curved arrows, Bezier connector lines, arc flows, curved timelines, circular migration paths;
• Wave diagrams, curved divider zones, irregular masks, complex decorative curves, branded lines;
• Map boundaries, contour lines, non-rectangular regions, complex icon contours, custom rings or non-standard arc KPI;
• The user explicitly states "1:1," "no offset allowed," "the curve amplitude is wrong," "streamlined," "restore according to the image," or "do not rectangularize."

After triggering, the following must be produced:

• Precisely cropped reference region image path;
• geometry_analysis geometry decomposition record, explaining whether the target is a stroke or fill shape, single line/double line/flow ribbon/wedge/area band/irregular boundary, endpoints, maximum curvature points, width changes, local protrusion/indentation direction, overlap hierarchy, and reconstruction method;
• The coordinates of the cropped image in the blueprint, the final PPT insertion frame coordinates, and the scale relationship;
• Boundary point, control point, endpoint, or key inflection point sampling record;
• Trace debug image, showing sampling or overlay inspection;
• Redrawn asset path for SVG path or PPT custom geometry;
• Local cropped image or local comparison image path of the same region after rendering;
• The corresponding component's trace_required: true、trace_method、trace_reference_crop、geometry_analysis、trace_debug_artifact、rendered_crop_comparison and asset path in the manifest.

Precise Tracing Execution Order Gate

After triggering the precise tracing gate, execution must follow the order below; if any step is missing, the final PPTX must not be generated, and curve_fidelity_pass must not be written as true：

1. Crop the target local reference image; the full-page blueprint must not be used as a substitute.
2. Write geometry_analysis, first determine the target geometry type and visual semantics, then decide whether to use stroke, fill path, polyline, PPT freeform/custom geometry, or tightly cropped non-text image.
3. Sample boundary points, control points, endpoints, maximum curvature points, and local features, and generate trace debug or overlay.
4. Draw SVG/custom geometry based on the geometry decomposition and sampling results.
5. Place into PPTX and render the page.
6. Crop the rendered local image of the same region and compare it with the reference local image.
7. Only after the local comparison passes is it allowed to curve_fidelity_pass=true。

Hard requirements:

• If geometry_analysis is not completed, the final SVG must not be drawn; you must not first draw the SVG by feel and then add the description.
• When the original image is a filled irregular shape, flow band, area band, or a boundary with varying width, you must not pass it off using a uniform-width stroke.
• When the original image is a double boundary/flow band, you must trace the upper and lower boundaries separately; you must not draw only the centerline.
• When the original image has local protrusions, recesses, spikes, gaps, or overlapping layers, you must record the direction and position; drawing the direction reversed is a failure.
• Using SVG/custom geometry only indicates that a controllable geometry tool was chosen; it does not indicate that precise tracing passed. An SVG without crop images, geometry decomposition, debug, and local rendering comparison is still considered a rough redraw.
• The whole page looking close cannot replace local comparison. If the triggered tracing region has no rendered_crop_comparison, delivery confirmation is not allowed.

Hard requirements:

• You must not, in order to preserve pictures=0, change a curve-semantics chart into incorrect native rectangles, parallelograms, hard-edged bar charts, or ordinary stacked bars.
• When curves, flow bands, or irregular boundaries are the core semantics of the chart, visual semantics take priority over pictures=0. Small-area, text-free SVG path assets are allowed, but must be registered in the manifest.
• SVG or image assets must not carry primary text, numbers, labels, sources, footers, or SO WHAT; these must still be rebuilt using PowerPoint native text.
• ImageGen must not be used for 1:1 geometric restoration. It can only be used for style exploration or non-precise illustration, and cannot replace pixel tracing, SVG path, or custom geometry.
• If SVG/custom geometry still cannot achieve the blueprint geometry, you must first explain the deviation and obtain user approval before using tightly cropped images of non-text visual regions; you must not use full-page or large-area blueprint screenshots.
• If the precise tracing gate is triggered but there is no trace crop, trace debug, tracing method, or manifest record, it is considered a process failure, and you must not proceed to page-by-page confirmation.
• After the user points out a deviation in curves, flow bands, arcs, or irregular boundaries, you must immediately mark that page's visual_qa_gate.json as curve_fidelity_pass=false and deliverable_allowed=false, and re-execute the local comparison process for the region pointed out by the user.

Label avoidance and curve quality gate

When a page contains a central diagram, flowchart, architecture diagram, ecosystem diagram, matrix diagram, timeline, path diagram, or any icon/node/curve-dense region, a label avoidance check must be performed. This check cannot be replaced by strict QA, image count, editability, or font size checks.

Hard requirements:

• Primary text, chart labels, values, legends, sources, SO WHAT, and body text must not overlap with icons, nodes, arrows, curves, rings, borders, data lines, or key shapes.
• Labels are allowed to be placed inside color blocks, bars, or nodes only when this is the blueprint's explicit design language and is registered in the manifest or visual QA as allowed_text_overlaps; unregistered overlap is always treated as a failure.
• When using native geometry to rebuild complex diagrams, you must record in the manifest or visual_qa_gate.json the label_collision_check. Missing it is a failure.
• If any main diagram label visibly overlaps with an icon/node/curve/ring/arrow, it is considered High/Critical, and delivery confirmation is not allowed.

Spatial anchor and relative position restoration gate

When a page contains a central diagram, flowchart, architecture diagram, ecosystem diagram, matrix diagram, timeline, path diagram, icon-dense diagram, or any combination of "icons + nodes + labels + arrows/curves," a spatial anchor check must be performed. This check cannot be replaced by label avoidance, curve tracing, font size checking, or the pictures count.

Before production, the blueprint's anchor relationships must first be recorded:

• The reference box, center point, width-to-height ratio, and relative order of each node/ring/card/icon;
• The offset of icons relative to node centers, the offset of labels relative to icons/nodes, and the offset of body items relative to labels;
• The start point, end point, direction of arrows or connecting lines, and their spacing from node boundaries;
• The horizontal/vertical spacing, alignment baselines, grouping boundaries, and reading order of elements in the same group;
• The allowed deviation thresholds. By default, center point deviation must not exceed 2% of the blueprint's local width or height, and key nodes/icons/text labels/arrow endpoints must not exceed an equivalent deviation of 6px; when the user requires "1:1" or "no offset," it must not exceed an equivalent deviation of 3px.

The manifest must include spatial_registration_check, recording at least:

{
  "passed": true,
  "reference_crop": "analysis/region_reference.png",
  "rendered_crop": "analysis/region_rendered.png",
  "checked_groups": [
    {
      "id": "conversion_node_01",
      "items": ["node_circle", "icon", "node_label", "bullets", "incoming_arrow", "outgoing_arrow"],
      "anchor_rule": "icon centered in node; label below icon; bullets below node; arrows connect node midline",
      "anchor_points": [
        {
          "item": "node_circle",
          "anchor": "center",
          "blueprint_bbox_px": {"x": 420, "y": 310, "w": 86, "h": 86},
          "render_bbox_px": {"x": 423, "y": 312, "w": 84, "h": 85},
          "delta_px": {"x": 3, "y": 2, "w": 2, "h": 1},
          "tolerance_px": 6,
          "status": "passed"
        },
        {
          "item": "node_label",
          "anchor": "text_baseline_center",
          "blueprint_bbox_px": {"x": 386, "y": 404, "w": 154, "h": 24},
          "render_bbox_px": {"x": 388, "y": 405, "w": 153, "h": 24},
          "delta_px": {"x": 2, "y": 1, "w": 1, "h": 0},
          "tolerance_px": 4,
          "status": "passed"
        },
        {
          "item": "outgoing_arrow",
          "anchor": "start_end",
          "blueprint_bbox_px": {"x": 506, "y": 350, "w": 198, "h": 18},
          "render_bbox_px": {"x": 508, "y": 351, "w": 198, "h": 18},
          "delta_px": {"x": 2, "y": 1, "w": 0, "h": 0},
          "tolerance_px": 6,
          "status": "passed"
        }
      ],
      "max_center_deviation_px": 6,
      "max_label_baseline_deviation_px": 4,
      "status": "passed"
    }
  ],
  "failures": []
}

Hard criteria:

• If an icon does not land near the blueprint anchor point of the corresponding node/ring/card, it fails even if it does not overlap.
• anchor_points only writes top_left、center、baseline and other abstract anchor points, without writing blueprint_bbox_px、render_bbox_px、delta_px and tolerance_px, it is considered an incomplete spatial anchor check.
• If any key anchor point status != passed, the page must not be delivered for confirmation.
• If any key anchor point delta_px exceeds tolerance_px, do not write spatial_registration_pass as true。
• If labels, body items, or values drift noticeably relative to icons/nodes, are misaligned, too low, too high, displaced, or change reading order, it fails even if there is no overlap.
• Node titles, axis labels, chart labels, or body items must be checked item by item for the text box center and the first-line/key-line baseline; registering only a whole-group check without registering specific text anchors is considered a failure.
• If arrow/connecting line endpoints do not connect to the corresponding node boundary or centerline in the blueprint, or the direction or spacing is noticeably different, it fails even if the lines are editable.
• If elements in the same group are pulled apart, compressed, displaced, or locally shifted as a whole, causing the information group relationships of the original image to change, it is considered High/Critical.
• checked_groups[].status can only be passed or failed; ambiguous states such as mostly_passed、passed_with_tolerance、minor_issue、acceptable must not be used. Any occurrence of a non- passed state is a failure.
• After the user points out that text, icons, nodes, or arrows are "offset," "not in their original position," or "different from the original image position," the page must immediately be visual_qa_gate.json with spatial_registration_pass=false and deliverable_allowed=false, and the local image must be re-cropped and compared against the anchor points.

Container Boundary, Continuous Text, and Table Semantics Gate

In the third stage, the owning container of each text segment must be recorded according to the blueprint. Owning containers include cards, panels, chart areas, table cells, conclusion bars, SO WHAT, title areas, footer areas, icon nodes, process nodes, KPI boxes, and annotation boxes. Whether text is acceptable is judged by the owning container, not by the entire page canvas.

Hard criteria:

• Text, key numbers, bullet points, chart labels, and source notes must not cross the boundaries of their owning container, separator lines, table cells, conclusion bars, or SO WHAT boxes. Even if they do not exceed the page canvas, they fail as long as they cross the owning container.
• All owning containers must retain visible inner padding. The minimum safe distance between body text and container boundaries, dividers, icons, arrows, or chart elements must be recorded according to the blueprint anchors; if not recorded, it fails.
• Container overflow must not be solved by reducing the font size. The container size, line width, line breaks, partitions, or refined text must be adjusted until the Typography Scale, container boundaries, and visual density all pass simultaneously.
• A semantically continuous sentence, SO WHAT main clause, conclusion sentence, or table cell body text must not be split into multiple independent text boxes resulting in abnormal spaces, broken sentences, drift, or misaligned baselines. When local bolding or highlighting is needed, rich text within the same text box must be used; if splitting is unavoidable,continuous_text_flow_check the baseline, letter spacing, spaces, and reading order after splitting must be proven item by item to be consistent with the blueprint.
• Table text must first be classified by semantics, then mapped to the Typography Scale. Table body text, action items, risk items, explanatory sentences, recommendation sentences, long bullet points, and complete short sentences must be registered as T7 or T10; they must not be registered as T11 micro-chart labels.T11 It is only allowed for axis labels, ticks, legends, extremely short table labels, units, short column labels, or non-sentence micro-labels.
• Table font size must be subordinate to table area and semantic role. In large-area management tables, action matrices, and risk matrices, if cells carry complete judgments or action descriptions, the readable range of T7/T10 must be used, and achieved by rearranging column widths, row heights, line breaks, and content refinement; complete sentences must not be compressed into T11。
• Table density must be consistent with the blueprint. If the table body text is too small, causing large areas of blank cells, collapse of the reading center of gravity, or the page appearing empty, it is also considered table_density_pass=false。
• After the user points out text overflow, abnormal spaces, table text too small, empty layout, hollow cells, or semantic text being treated as micro-labels, the corresponding field of visual_qa_gate.json on that page must be immediately set to false, and deliverable_allowed=false, and after rework, re-render for comparison.

The manifest must record as needed:

{
  "container_overflow_check": {
    "passed": true,
    "checked_regions": [
      {"id": "so_what_body", "owner_container": "so_what_box", "status": "passed"}
    ],
    "failures": []
  },
  "continuous_text_flow_check": {
    "passed": true,
    "checked_text_runs": [
      {"id": "so_what_sentence", "method": "single-textbox-rich-text", "status": "passed"}
    ],
    "failures": []
  },
  "table_text_objects": [
    {"id": "risk_row_01", "semantic_role": "table_body", "role": "T7", "font_size_pt": 9.5}
  ],
  "table_density_check": {
    "passed": true,
    "checked_cells": [
      {"id": "action_table_r1c3", "semantic_role": "table_action", "status": "passed"}
    ],
    "failures": []
  }
}

Hard criteria for curve quality:

• After the precise tracing gate is triggered, a polyline of 5-6 points must not be used to impersonate a smooth curve.
• If curves, arcs, flow ribbons, or irregular boundaries are core visual semantics, SVG path, PPT freeform/custom geometry, or a sufficiently dense polyline must be used; by default, the number of sampling points for core curves must not be fewer than 16.
• If a polyline is used to approximate a curve, the manifest must record each core curve's point_count、min_required_point_count and max_deviation_px or clearly annotate the manual override inspection result.
• trace_debug_artifact Must display curve coverage, sampling points, or boundary comparison; drawing only a rough outer frame is unacceptable.
• If the curve's bending amplitude, endpoints, width changes, overlap order, or visual center of gravity clearly deviate from the blueprint, it must be reworked even if structural QA passes.

Final visual QA gate

Before each delivery confirmation in phase three, visual_qa_gate.jsonmust be generated. Without this file, delivery confirmation is not allowed. This file must give a clear boolean judgment page by page, and vague expressions such as "basically", "roughly", or "looks okay" must not be used.

visual_qa_gate.json is not a self-declaration. Each true must be bound to an evidence record; without evidence, truemust not be written. Before deliverable_allowed=true on each page, at least the following must be provided:

• Approved blueprint image:blueprint_render_path；
• Current PPT rendering image:ppt_render_path；
• Side-by-side comparison image of the blueprint and the PPT rendering:side_by_side_comparison_path；
• Local overlay / bbox comparison image of key regions:local_overlay_artifacts；
• Blueprint measurement table or coordinate conversion check file:measurement_evidence_path；
• Numeric anchor check file:spatial_numeric_check_path；
• Item-by-item visual difference record:visual_differences；
• Supporting each visual QA field's evidence mapping; string evidence must be an existing file path, and inline check records must be written as structured objects.

Each page must contain at least:

{
  "slide": 4,
  "blueprint_render_path": "blueprints/slide-04.png",
  "ppt_render_path": "renders/slide-04.png",
  "side_by_side_comparison_path": "qa/slide-04-side-by-side.png",
  "local_overlay_artifacts": [
    "qa/slide-04-title-overlay.png",
    "qa/slide-04-main-chart-overlay.png",
    "qa/slide-04-so-what-overlay.png",
    "qa/slide-04-footer-overlay.png"
  ],
  "measurement_evidence_path": "qa/slide-04-blueprint-measurement.json",
  "spatial_numeric_check_path": "qa/slide-04-spatial-numeric-check.json",
  "visual_differences": [],
  "evidence": {
    "surface_system_match": "qa/slide-04-color-samples.json",
    "main_chart_semantics_match": "qa/slide-04-chart-semantics.json",
    "visual_semantics_preserved": "qa/slide-04-side-by-side.png",
    "editable_information_layer_pass": "slide_manifest.json",
    "spatial_registration_pass": "qa/slide-04-anchor-check.json",
    "curve_fidelity_pass": "qa/slide-04-curve-check.json",
    "label_collision_pass": "qa/slide-04-label-check.json",
    "text_overflow_pass": "qa/slide-04-text-check.json",
    "container_overflow_pass": "qa/slide-04-container-check.json",
    "continuous_text_flow_pass": "qa/slide-04-text-flow.json",
    "table_semantic_typography_pass": "qa/slide-04-table-type.json",
    "table_density_pass": "qa/slide-04-table-density.json",
    "blueprint_background_not_used": "slide_manifest.json"
  },
  "surface_system_match": true,
  "main_chart_semantics_match": true,
  "visual_semantics_preserved": true,
  "editable_information_layer_pass": true,
  "spatial_registration_pass": true,
  "curve_fidelity_pass": true,
  "label_collision_pass": true,
  "text_overflow_pass": true,
  "container_overflow_pass": true,
  "continuous_text_flow_pass": true,
  "table_semantic_typography_pass": true,
  "table_density_pass": true,
  "blueprint_background_not_used": true,
  "deliverable_allowed": true,
  "notes": []
}

Hard criteria:

• If any key item is false，deliverable_allowed it must be false, and delivery confirmation must not be given.
• deliverable_allowed=true but missing blueprint_render_path、ppt_render_path、side_by_side_comparison_path or visual_differences, it is a failure.
• deliverable_allowed=true but missing local_overlay_artifacts、measurement_evidence_path or spatial_numeric_check_path, it is a failure.
• visual_differences contains High/Critical differences that the user has not explicitly accepted, delivery confirmation must not be given.
• If any visual field is true but lacks the corresponding evidence, it is a failure.
• strict QA passing cannot replace visual_qa_gate.json; structural compliance does not equal visual qualification.
• editable_information_layer_pass=true only proves that the information layer is editable, and does not mean the visual semantics are qualified;visual_semantics_preserved=true only proves that the visual semantics have not been degraded, and does not mean the information layer is editable. Both must simultaneously be true。
• spatial_registration_pass=true must be judged based on local cropped images or an anchor checklist; do not substitute with "no overlap" or "looks about the same".
• The final reply must not only say that QA passed; it must also state whether visual QA passed.
• If visual QA fails, rework must continue or it must be explicitly marked as "not passed, cannot confirm"; a failed page must not be treated as a qualified delivery.

Misinterpretations and corrections

Misinterpretation	Correct approach
"For visual fidelity, first lay down the full-page blueprint and then cover the text"	Wrong. The blueprint is a reference, not a delivery background. Simple charts and main text must be natively rebuilt.
"The chart looks complex, so it can be kept as an image"	Wrong. Line charts, bar charts, axes, comparison bars, labels, and CAGR/KPI are by default simple charts and should be natively rebuilt.
"As long as the main title is editable, it is fine"	Wrong. Body text, chart labels, key numbers, notes, footers, page numbers, and SO WHAT must also be editable.
"The user asked to reproduce according to the blueprint, so a blueprint screenshot can be used"	Wrong. Reproducing according to the blueprint means reproducing the layout, hierarchy, density, and visual language; it does not mean putting the blueprint screenshot into the PPT.
"The font size is too small, but all the content fits"	Wrong. Do not replace structural rearrangement with shrinking the font size; prioritize condensing text, adjusting sections, or splitting pages.
"To keep pictures=0, change flow/curve diagrams into rectangles or parallelograms"	Wrong. Curve semantic diagrams must first be precisely traced; when necessary, a small area of text-free SVG is allowed and must be registered in the manifest.
"I ensured editability, so I can simplify the blueprint visuals"	Wrong. Editability and visual semantics are equally hard thresholds; use a hybrid approach of native objects plus small-area text-free SVG/custom geometry.
"It looks visually similar, so it is fine to make the main information into an image"	Wrong. Visual fidelity cannot override information editability; main text, numbers, labels, SO WHAT, and footers must remain natively editable.
"Strict QA passed, so the visual reproduction is acceptable"	Wrong. Strict QA only proves that some structural rules passed; the background system, panel logic, chart semantics, and curve geometry still require rendered comparison.
"ImageGen can draw it more accurately, so use it to redo the precise curves"	Wrong. 1:1 geometric reproduction cannot rely on random generation; it must use cropping, sampling, SVG paths, or custom geometry.
"SVG is also an image, so why not just crop the blueprint directly"	Wrong. SVG paths are controllable geometric assets; a screenshot of the blueprint is a non-editable background shortcut. Main text and labels must still be native.
"The text and icons overlap a bit, but structural QA passed"	Wrong. Visible overlap between labels, values, body text and icons/nodes/curves/arrows is a visual QA failure and must be reworked.
"The text and icons do not overlap, so the positions are acceptable"	Wrong. No overlap only passes label avoidance; icons, labels, nodes, and arrows must also pass the spatial anchor/relative position reproduction gate.
"The overall layout is about right, a slight local icon offset does not matter"	Wrong. Center diagrams, flowcharts, ecosystem diagrams, and path diagrams depend on local anchor relationships; noticeable drift of icons, labels, nodes, or arrows is a failure.
"The text does not go outside the page, so it does not count as overflow"	Wrong. Text must stay within its owning container; crossing the boundary of a card, cell, SO WHAT, conclusion bar, or chart area is a failure.
"Splitting one sentence into multiple text boxes makes highlighting easier"	Wrong. Continuous semantic text must remain a continuous flow; if splitting causes abnormal spaces, broken sentences, baseline misalignment, or drift, it is a failure.
"Table text can all use T11 by default"	Wrong. Table body text, action items, risk items, and complete judgment sentences must be registered as T7/T10 by semantics; T11 is only for micro labels.
"The table font size is not below the lower limit, so empty layout is fine"	Wrong. Table density must match the blueprint; large blank areas, collapsed reading focus, or cells that look empty are failures.
"Connecting a few points can also express the meaning of the curve"	Wrong. Core curves must not be faked with a few polyline points; use path/freeform/custom geometry or sufficiently dense sampling.
"I have already provided the rendered image, the user will see the problems themselves"	Wrong. Before delivery, the visual QA gate must be proactively filled in; any failure found must block delivery and must not be shifted to the user to find errors.

Text reconstruction requirements

1. The original image's font character, font size hierarchy, font weight, color, line spacing, letter spacing, alignment, and position must be reproduced.
2. If the font cannot be identified, choose the system font closest to the visual character of the original image.
3. For Chinese, prefer styles close to the original image such as Heiti/PingFang/Source Han Sans/Microsoft YaHei; for English and numbers, choose sans-serif or serif fonts close to the original image.
4. Obvious typos, missing characters, incorrect line breaks, text overflow, text overlap, or positional drift are not allowed.
5. Keyword highlighting, differently colored text, weight variations, numbering, quotation marks, units, and punctuation must all be preserved.
6. If the layout is followed exactly as in the original image, the native PPT text will look noticeably worse; slight layout adjustments may be made without changing the meaning.
7. Body text must not use an excessively small font size to gain density; a fixed Typography Scale must be used. If space is insufficient, prioritize refining the text, reorganizing the layout, or splitting the information rather than continuing to shrink the font.

Fixed Typography Scale

A fixed set of 15 text levels throughout:C0 is dedicated to covers/section dividers,T1-T14 are content page levels. Both production and QA must check font size, weight, and readability according to these names.

Level	Name	Typical Position	Font Size Range
C0	Cover/Section Divider Title	Cover main title, section divider title	32-44pt
T1	Page Number/Section Badge	Top-left page number badge, section number badge	14-18pt
T2	Page Main Title/Conclusion Title	Conclusion sentence at the top of each page	22-28pt
T3	Page Subtitle/Context Description	Subtitle/kicker below the main title	10-12pt
T4	Module Title/Chart Title/Information Area Title	Chart title, right panel title, information area title	11-14pt
T5	Evidence Number/Lightweight Label	E02/E05 labels, status labels, small badges	7.5-8.5pt
T6	Evidence Block Title/Subsection Title	Left-side evidence block title, paragraph subheading	11-13pt
T7	Body Explanatory Paragraph	Evidence explanation, management interpretation body text	9.5-11pt
T8	Conclusion Bar Text/Core Conclusion	Blue conclusion bar, core summary box	10-12pt
T9	SO WHAT Label	SO WHAT / WHY IT MATTERS label	10-12pt
T10	SO WHAT Body/Business Implication	Bottom SO WHAT body text, action implications	9.5-11pt
T11	Chart Axis/Legend/Tick/Microchart Label	Coordinate axis, legend, tick, microchart label	7.5-9pt
T12	Chart Data Label/Direct Annotation	Line point values, bar labels, percentage annotations	8.5-11pt
T13	Key KPI/Large Number	4.2%, CAGR, large numbers for key metrics	18-28pt
T14	Annotation/Definition/Source/Footer	Annotations, caveats, sources, footers, small page numbers	6.5-8pt

Mapping of 1-8 in the user's image: 1=T2, 2=T3, 3=T6/T4, 4=T7, 5=T14, 6=T11/T12/T13, 7=T10, 8=T14.

Design Restoration Requirements

1. Maintain the original image's layout, proportions, visual focus, reading order, and whitespace rhythm.
2. Maintain the original image's colors, contrast, light-dark relationships, light source direction, shadow intensity, and spatial hierarchy.
3. Maintain the original texture of complex visual areas; do not replace them with default PPT icons, default charts, default shadows, or templated cards.
4. The position, size, hierarchy, cropping method, corner radius, stroke, shadow, transparency, and alignment relationships of all elements must closely match the original image; exceeding the recorded tolerance is a failure.
5. If the page has charts, flowcharts, matrices, timelines, comparison charts, cycle diagrams, or infographics, their visual relationships and reading logic must be preserved.
6. If a non-information-layer visual element can be edited and would significantly reduce the perceived quality, it must first be registered as a complex visual asset with the reason stated; primary text, numbers, labels, footers, and SO WHAT must not be turned into images.
7. The page must follow the unified page surface system of the chosen style; each module must not be arbitrarily changed into large white cards; it must be restored according to the background color, color levels, borders, column headers, and divider lines recorded in the blueprint.
8. If the semantics of the main chart depend on curves, stream bands, arcs, or irregular boundaries, it must be restored according to the precise tracing gate and must not be changed into a basic chart with different visual semantics.

Output and Check

1. Output one PPTX file.
2. At the same time, render a PPTX preview image and make a visual comparison with the original image.
3. If there are unrecorded deviations between the preview image and the original image in texture, proportions, spacing, color, hierarchy, container boundaries, table density, or overall look and feel, continue iterating and revising.
4. Finally, briefly explain:
   • which elements are editable text/shapes/charts;
   • which elements are image assets cropped or extracted from the original image;
   • which text is kept as images because it is embedded in complex visual areas;
   • where editability was sacrificed to preserve visual texture.

Non-negotiable controls

• All facts and numbers must come from the source materials. Do not fabricate missing evidence.
• Record inconsistent values or calculation bases between sources; do not silently normalize them.
• Phase one must include a content brainstorming substep: based on the evidence table, propose 2-3 optional storylines, build an issue tree or hypothesis tree, compare each storyline's core conclusion, evidence strength, risk gaps, and suitable audience, then recommend one path.
• The first confirmation cannot confirm only the page titles. It must also confirm the storyline, page count, page-by-page arguments, page information density, and component inventory, including how many information zones per page, the ratio of the main chart to the sidebar, the number of tables/annotations/legends/micro-charts, key evidence IDs, and the SO WHAT.
• Page information density and component inventory are inputs to the page-by-page blueprint in phase two, and must not be downgraded again during the blueprint stage into loose cards or sparse large-type posters.
• Use SCR throughout, and make each content page title a defensible conclusion. Titles must withstand MBB-style challenge: what evidence supports it, what caveat might overturn it, and what decision does it change?
• By default, display 8 fixed CyberPPT visual styles. One may be recommended based on the material type, but ultimately let the user choose.
• Visual samples must be independent, complete 16:9 pages; collages, thumbnail walls, or compressed multi-page canvases must not be used.
• Before confirming the style samples, 8 independent 16:9 images must be sent directly through the current conversation. Prefer using the built-in samples assets/palette-samples/palette-01.png to palette-08.png, or generate 8 new real images.
• Do not deliver only Markdown, text lists, tables, style descriptions, web pages, HTML, URLs, folder paths, file lists, or stage2_style_options.md as the phase-two deliverable; these can only serve as auxiliary explanations beside the images and cannot replace the 8 independent sample images in the current conversation.
• If a web page is used as an aid, the web page can only be an additional browsing method; the web page must not be used as the sole basis for style confirmation.
• Using only source files does not equal skipping visual samples; "using only source files" only constrains the source of facts, data, text, and conclusions, and does not cancel the fixed display of 8 visual samples.
• Before generating default-style samples, you must explicitly list and use each of visual-system.md the fixed 8 names and palettes; "8 visual directions" does not equal the fixed 8 CyberPPT styles.
• After the visual system is confirmed, you must first generate a page-by-page ImageGen blueprint for the complete page count, then proceed to hybrid restoration PPTX. The blueprint is only used for composition, hierarchy, density, and chart language.
• The page-by-page blueprint stage must continue to use the selected style and must not diverge again; each page's blueprint prompt and inspection record must include the locked style number, palette, grid, information density, and chart language.
• Unless the user explicitly requests skipping ImageGen, the page-by-page blueprint must be generated by ImageGen as complete 16:9 bitmap images; PptxGenJS, python-pptx, HTML, CSS, SVG, canvas, Pillow, matplotlib, PowerPoint, or any local drawing script must not be used to directly draw the blueprint.
• Scripts may organize prompts, save/copy/rename ImageGen output images, generate metadata, manifests, QA reports, comparison images, or inspection overlays; but they cannot serve as the final image generator for the second-stage page-by-page blueprint.
• Skip ImageGen only when the user explicitly requests it. When skipping, record the reason and use the template, reference images, or visual specifications provided by the user.
• Treat ImageGen text and numbers as unreliable. Generated images are only used as references for visual composition and art direction.
• The third stage must simultaneously uphold the two hard thresholds of an editable information layer and high-fidelity visual semantics; neither goal may be allowed to override the other.
• When restoring PPTX, you must first distinguish the complex visual asset layer from the editable information layer; primary text must be rebuilt using independent editable PowerPoint text boxes.
• Do not use full-page screenshots as a shortcut. The final primary text must be editable and must not be baked entirely into images.
• Complex visual areas may be cropped into images to preserve texture, but you must state which content thereby sacrifices editability.
• Do not stop at the chart skeleton. Every content page must include conclusions, evidence, interpretation, and business implications or SO WHAT.
• The default consulting-style cover remains low-density: title, subtitle, version/date/context are sufficient; unless the user explicitly requests it, do not add KPI cards or charts.
• Lock the page size before layout. The coordinate system, visual references, rendered images, and output files must use the same aspect ratio.
• Every page must be rendered and inspected. File generation, successful compilation, or valid XML does not equal acceptable visual quality.

Production and Verification

Use the most appropriate PPT tool in the current environment. Preserve the original source files and output version-numbered files.

Run the structure validator:

python scripts/validate_pptx.py path/to/deck.pptx --manifest path/to/slide_manifest.json --visual-qa path/to/visual_qa_gate.json --strict --json-out path/to/report.json

In the third stage,--manifest --visual-qa --strict cannot be omitted. In strict mode, the appearance of hard-rule errors means failure, and it must be reworked, then regenerated and revalidated; it must not merely be explained away or ignored.

Treat non-hard-rule validation warnings as review prompts, not absolute visual judgments. Check each rendered page for blank-space imbalance, overly small font sizes, cropping, overlap, chart label collisions, weakened hierarchy, and style drift.

Before page-by-page confirmation in the third stage, you must simultaneously provide:

• PPTX path;
• current page single-page PPTX path;
• full-page rendered image;
• Approved blueprint image, PPT rendered image, and side-by-side comparison image;
• Local comparison image of key areas on the current page;
• QA JSON；
• visual_qa_gate.json Path and item-by-item results;
• slide_manifest.json Path;
• page_execution Summary, confirming mode=single_page、page_status=approved、user_confirmed=true and made_before_next_slide=true；
• pictures Quantity;
• Count of native text, shapes, charts, and tables;
• List of image assets on this page and the reason for retaining each item;
• Which text, numbers, chart labels, footers, and SO WHAT on this page are editable;
• If this page is a simple chart page, explain why pictures is 0, or explain the reason for the exception.
• If this page triggers the curve/irregular visual precise tracing gate, provide trace crop, trace debug, tracing method, SVG/custom geometry assets, and a summary of manifest fields.
• If this page contains a central diagram/flowchart/architecture diagram/ecosystem diagram/matrix diagram, provide label avoidance check results; if any overlap is unresolved, confirmation must not be given.
• If this page contains dense areas of icons, nodes, labels, arrows, or curves, provide spatial anchor check results; if any offset is unresolved, confirmation must not be given.
• If this page contains cards, panels, tables, conclusion bars, SO WHAT, chart annotations, or fixed-area text, provide container boundary check results; if any text crosses beyond its owning container, confirmation must not be given.
• If this page contains split text, rich text highlights, or continuous sentences spanning regions, provide continuous text flow check results; if there is any abnormal space, sentence break, baseline misalignment, or drift, confirmation must not be given.
• If this page contains tables, matrices, action lists, risk lists, or grid-based management tables, provide table semantic font size and table density check results; if table body text is registered as T11 or cells are largely empty, confirmation must not be given.
• Every visual_qa_gate.json in which is true visual field must provide evidence. String evidence must be an existing file path, and inline check records must be written as structured objects; without evidence, delivery confirmation cannot be given.

Common failures

• Making style samples into a collage to save time will undermine the 16:9 judgment.
• Outputting only stage2_style_options.md or a text style list, without showing the 8 actual images, will cause Phase 2 to be non-compliant.
• Providing only a webpage, HTML, URL, folder path, or file list, without directly sending the 8 sample images through the current conversation, will cause Phase 2 to be non-compliant.
• If images on the webpage fail to load, paths are invalid, or the user cannot see the samples, yet you continue to ask the user to choose a style, it will cause Phase 2 to be non-compliant.
• If the user has not yet confirmed seeing the 8 samples and you proceed to the page-by-page blueprint phase, it will cause Phase 2 to be non-compliant.
• If the user did not request English and English is not a valid target language, yet you generate an English blueprint by default, it is a Phase 2 failure.
• If the main visible text language of the blueprint is inconsistent with the valid target language, it is a Phase 2 failure.
• If each page's blueprint prompt does not record target_language、language_source and effective_language, it is a Phase 2 failure.
• If there is page-level, section-level, or component-level language override but language_overridesis not recorded, it is a Phase 2 failure.
• If the user only requested that a certain page use English, but other uncovered pages are also changed to English, it is a Phase 2 failure.
• If the blueprint or final PPT contains target_language、language_source、effective_language、language_overrides、allowed_foreign_terms, "target language=Chinese", or language=Chinese Waiting for execution metadata visible text, which is generated text contamination and must be redone.
• Starting execution without reading the stage reference will miss fixed style, confirmation gates, blueprint locking, or hybrid restoration constraints.
• Skipping content brainstorming in the first stage and only providing a single-version outline will lead to insufficient information density in the subsequent blueprint.
• A page-by-page outline without page information density and component inventory will cause ImageGen to default to generating loose cards, generalized charts, and low-density layouts.
• Using generate_stage2_blueprints.py, PptxGenJS, python-pptx, HTML, CSS, SVG, canvas, Pillow, matplotlib, or PowerPoint to directly generate page-by-page blueprints is a second-stage failure; these tools cannot replace ImageGen.
• Page-by-page blueprints that look like ordinary PPT card pages, HTML dashboards, wireframes, structural sketches, script drawings, or low-fidelity mockups are a second-stage failure.
• Downgrading ImageGen blueprints to regular placeholder images, default card pages, or script sketches for the convenience of third-stage measurement is a second-stage failure.
• Only increasing the number of cards or the amount of text without restoring the confirmed visual system, hero image, evidence area, annotations, sidebar, SO WHAT, caveat, evidence ID, and chart language still counts as a second-stage failure.
• Replacing the fixed 8 CyberPPT styles with improvised "8 directions" will make the second step non-compliant.
• Diverging again in the page-by-page blueprint after the style is selected will cause style drift.
• Reconnecting the old element-level image pipeline will conflict with the current third-step hybrid restoration strategy.
• Reducing font size instead of restructuring is not an acceptable solution.
• Only making charts without explanatory text will fail to reach the information density of a consulting report.
• Matching only colors is not equivalent to reproducing the visual system; grid, density, hierarchy, chart language, and whitespace behavior must also be matched.
• Copying labels from generated images will contaminate the evidence chain.
• Treating blueprint screenshots as final pages will break editability.
• Laying a full-page blueprint background on the grounds of "restoring according to the blueprint" is a wrong interpretation; restoring according to the blueprint means reproducing the composition, hierarchy, density, and visual language, not delivering with blueprint screenshots.
• Keeping line charts, bar charts, axes, comparison bars, tables, or SO WHAT as images as complex visual assets will break editability; these must by default be rebuilt natively.
• pictures > 0 However, the page component inventory states "no image assets need to be retained," which is a third-stage rework item and cannot be delivered for confirmation.
• Baking the main text into images violates the process. Small text inside complex visual areas may be retained as images only when splitting would damage the texture, and this must be explained.
• Rebuilding complex visual effects into low-quality default graphics for the sake of full editability will damage the premium feel and restoration fidelity.
• Rebuilding the unified page surface system into a stack of white cards will damage the visual system, and must be reworked even if the text and image counts are compliant.
• Changing streamline diagrams, Sankey diagrams, ribbon migration diagrams, or arc flow diagrams into hard-edged rectangles, parallelograms, or ordinary stacked bars is a chart semantic error; it must not be rationalized with pictures=0 .
• Triggering the curve/irregular visual precise tracing gate without cropped images, sampling records, trace debug, or manifest records is a process failure.
• Using ImageGen to try to fix 1:1 curve offsets is the wrong path; controllable tracing and redrawing must be used.
• Text labels covering icons, nodes, arrows, rings, or curves, even if font size and editability are compliant, still constitutes a visual QA failure.
• Text, icons, nodes, arrows, or connector lines clearly offset relative to blueprint anchors, even without overlap, still constitutes a visual QA failure.
• Text does not exceed the page but crosses the boundaries of its own card, panel, table cell, SO WHAT, conclusion bar, or chart area, which constitutes a container boundary failure.
• Continuous sentences producing abnormal spaces, line breaks, baseline misalignment, or drift due to split text boxes constitutes a continuous text flow failure.
• Table body text, action items, risk items, explanatory sentences, or recommendation sentences registered as T11 micro-labels constitutes a table semantic font size failure.
• Table content font size being too small, causing large empty areas in cells, collapse of reading focus, or a layout that looks empty, constitutes a table density failure.
• Using a few polyline points to piece together core arcs, flow lines, or irregular boundaries that should be smooth is a curve high-fidelity failure.
• Without visual_qa_gate.json, missing key visual QA items, or deliverable_allowed=false still delivered, constitutes a process failure.
• visual_qa_gate.json only writing boolean values, without blueprint images, PPT rendered images, side-by-side comparison images, or field evidence, constitutes a process failure.
• Missing blueprint_reconstruction_plan and directly generating PPTX constitutes a process failure.
• Generating the complete deck all at once in the third stage and treating it as the final version constitutes a process failure; the high-fidelity final version must be produced page by page and confirmed page by page.
• Generating the complete PPTX in batches first, then retroactively writing the manifest, visual QA, or side-by-side, constitutes fake page-by-page acceptance.
• When merging the final deck, regenerating pages, re-laying out, redrawing charts, reapplying backgrounds, or using single-page rendered images as full-page backgrounds constitutes a merge failure.
• After merging, not exporting PNGs for regression verification, or after merging showing background color changes, object offsets, font changes, missing images, SVG deformation, chart misalignment, or decreased information density relative to the already-passed single-page renders, constitutes a merge failure.
• Moving or stretching image assets during PPT reconstruction will damage blueprint fidelity.
• Applying 10 x 5.625 coordinates on a 13.333 x 7.5 canvas will cause abnormal blank space on the right or bottom.
• Checking only the first page will miss repeated layout, label, and density issues.