A Scientific Image and Text Human-Computer Collaborative Generation System and Method Based on Visual Mapping Verification

The human-computer collaborative generation system for scientific text and image, which uses visual mapping verification, solves the problems of accuracy of scientific content and consistency across pages in the generation of scientific text and image, and achieves efficient and controllable text and image generation, supporting the development of teachers' meta-task awareness and educational research.

CN123118045APending Publication Date: 2026-11-17HUANGGANG NORMAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611327636.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing scientific text and image generation technologies suffer from problems such as undetermined accuracy of scientific content, uncontrolled cross-page consistency, inability to terminate the correction process, and disordered human-computer collaboration framework. These issues make it difficult to guarantee the scientific accuracy of the generated text and image content, maintain cross-page consistency, render the correction process ineffective and uncontrollable, and make the human-computer collaboration trajectory unverifiable.

Method used

A human-computer collaborative generation system for scientific text and graphics based on visual mapping verification is adopted. Through modules such as knowledge initiation and carrier determination, visual mapping table generation and gating, meta-task status card persistence, page-level collaborative loop engine, verification and differentiation correction, and autonomy regulator, a hard gating system for visual mapping table, persistent data structure, multi-level verification and correction strategies, and multiple levels of human intervention density are established to ensure the accuracy of scientific content and consistency across pages.

Benefits of technology

It achieves accuracy and cross-page consistency in the generation of scientific graphics and text, supports the development of teachers' meta-task awareness, avoids infinite loops and hidden dependencies, improves generation efficiency and controllability, and provides auditable collaboration trajectories to support teacher training and educational research.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118045A_ABST
    Figure CN123118045A_ABST
Patent Text Reader

Abstract

This invention proposes a scientific graphic and text human-computer collaborative generation system and method based on visual mapping verification. The method includes the following steps: acquiring knowledge files and determining the carrier scheme; converting the knowledge content into a user-confirmed visual mapping table; persistently storing task-fixed attributes as a state data structure with version numbers; executing an iterative loop of generation and verification based on the state data; classifying the judgment results into multiple levels and executing corresponding correction strategies, terminating the loop after triggering a preset termination condition; setting multiple levels of human intervention density and driving level migration. This invention uses a visual mapping table as a hard gate to carry epistemological filtering of subject knowledge, uses meta-task state cards to persistently define the human-computer collaboration framework, and uses three-value verification and promotion / demotion rules to drive a terminateable generation-evaluation-correction loop. This allows for the efficient generation of scientifically accurate, consistent, and teacher-supporting popular science graphic and text publications in a non-programming environment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a scientific image and text human-computer collaborative generation system and method based on visual mapping verification. Background Technology

[0002] The combination of large language models and text-to-image technology allows teachers and researchers to transform knowledge documents such as papers and lecture notes into popular science publications for readers of different ages. However, the existing single-round generation paradigm of "prompt word engineering" suffers from the following systemic failures in this scenario: 1) Scientific illusion (the engine "fills in" errors in direction, proportion, and causal structure, and this is subtle to readers outside the field). 2) Long conversation state decay (character and art style locking statements drift with context distance); 3) Self-evaluation bias in validation (the same model is generated and then self-evaluated, lacking objective criteria); 4) The modified loop has no termination condition (repeatedly retrying the same strategy); 5) Manual confirmation bottleneck (interaction costs increase linearly with the number of pages).

[0003] At the level of educational theory, Wong (2025) points out that generative AI is moving from the "prompt era" to the "post-prompt era": the burden of human-computer interaction shifts from prompt construction to trajectory interpretation, and users need to develop meta-task awareness (MTA) – awareness of the deep goals, structure and evaluation criteria of tasks – and connect fragmented single-point interactions into learning chains (CoL) and instructional design evaluation chains (CoLDE); at the same time, it warns that the scaffolding generated by the system may cause "hidden dependence" and erode user subjectivity. Chen and Wong (2026) further propose the MTA-TPACK dynamic collaborative spiral: static TPACK knowledge resources (subject content knowledge CK, pedagogical knowledge PK, technical knowledge TK) must be activated by the "dynamic navigation engine" MTA in the Generate-Evaluate-Refine iterative cycle, and the generation process is constrained by "mechanism-visual indicator mapping" and anchored scoring criteria, ultimately producing Visible Pedagogical Thinking (VPT) – an expert reasoning trajectory that is indexed by criteria, auditable and reusable. Taking the scientific visualization of typhoon-terrain interaction as an example, this study empirically demonstrates that both the human-led (Midjourney four-cycle) and AI closed-loop self-assessment (GPT-4o four-cycle) paths can converge to expert-level accuracy (2 / 5→5 / 5), and records the four stages of AI self-assessment maturity (descriptive comparison → diagnostic assessment → structured introspection → selective stability).

[0004] The aforementioned theoretical work clarifies "how" human-machine collaboration "should be" (meta-task awareness, standard-anchored evaluation, evaluation feedback design, and auditable trajectory), but its experimental process relies on manual operation by researchers: the mapping table exists in the paper's table rather than the system gate, the evaluation standards rely on on-site expert citation, the cyclical advancement and upgrading rely on manual judgment, and the VPT trajectory is compiled afterward.

[0005] At the engineering level, harness engineering and loop engineering are expected to form methodologies around 2026, but their application will primarily focus on areas with compiler / automatic test verifier capabilities, such as large language model code generation. Translating the process model of educational theory into a deployable intelligent agent system with hard gating, state persistence, termination guarantees, and telemetry capabilities is a crucial technological gap that urgently needs to be filled in this field.

[0006] In summary, existing scientific image and text generation technologies still suffer from the following technical shortcomings: Mechanism-visual indicator mapping remains at the research methodology level, failing to be formalized into a systemic mandatory intermediary and gating mechanism, thus unable to prevent bypassing; meta-tasks (goals / constraints / evaluation criteria / human-machine division of labor) lack persistent data structures, relying on dialogue memory, inevitably leading to drift during long conversations and cross-conversations; evaluation relies on model subjective self-assessment or expert on-site judgment, lacking objective judgment on each item and differentiated correction strategies; correction loops lack termination conditions and strategy upgrade mechanisms; the redistribution of human-machine evaluation labor (human-led ↔ AI closed loop) lacks a controlled tiered mechanism and promotion / demotion rules, resulting in either entirely manual confirmation (inefficient) or unchecked automation (out of control and exacerbating implicit dependencies); collaborative trajectories are reconstructed post-event, distorted and unverifiable, unable to support teacher training and reuse in educational research.

[0007] References: [1] H. -C. Chen, (2023) Harnessing AI for Scientific Illustration: Exploring Tornado Dynamics Through Midjourney, 12th International Conference on Awareness Science and Technology (iCAST), 2023, pp. 136-141 [2] Wong, L.-H. (2025). Transforming Education with Generative AI: Designing for Post-Prompting Era. Information and Technology in Education andLearning, 5(1) [3] Chen, H.-C., & Wong, L.-H. (2026). The MTA-TPACK DynamicCollaboration Spiral: Making Pedagogical Thinking Visible in Human–AIScientific Visualization for Sustainable Teacher Innovation. Sustainability,18, 2718 Summary of the Invention

[0008] This invention proposes a human-computer collaborative generation system and method for scientific text and image based on visual mapping verification, which solves the problems of undetermined accuracy of scientific content, loss of consistency across pages, inability to terminate the correction process, and disordered human-computer collaboration frame definition in existing scientific text and image generation technologies.

[0009] The technical solution of this invention is implemented as follows:

[0010] The first aspect of this invention provides a scientific image and text human-computer collaborative generation system based on visual mapping verification, comprising:

[0011] The knowledge initiation and carrier determination module is used to acquire knowledge files and determine the carrier scheme for graphic and textual publications;

[0012] The visual mapping table generation and gating module is used to convert the knowledge content in the knowledge file into a visual mapping table. The visual mapping table can only enter the subsequent image generation process after the user confirms each item in the visual mapping table.

[0013] The meta-task status card persistence module is used to organize the fixed attributes and constraints of this image and text generation task into a status data structure for persistent storage.

[0014] The page-level collaborative loop engine is used to execute an iterative loop containing generation and verification for each page. The generation is anchored by a state data structure, and the verification is based on a visual mapping table to determine the generated content.

[0015] The verification and differentiation correction module is used to classify the judgment results into multiple levels, execute the corresponding level correction strategy according to the classification results, and terminate the correction loop after triggering the preset termination condition.

[0016] The autonomy regulator is used to set multiple levels of human intervention density. It drives the shift of levels based on objective indicators of loop operation, and the upgrade requires user authorization.

[0017] Specifically, the knowledge initiation and carrier determination module is configured as follows:

[0018] The system prioritizes reading knowledge files provided by the user; if no knowledge files are available, it automatically retrieves and summarizes the core mechanisms and submits them for user confirmation.

[0019] Based on a four-dimensional judgment table of reader level, knowledge structure, emotional goals and material form, output one primary and one backup carrier proposal including story picture book track and science magazine illustrated book track, with each carrier proposal accompanied by a recommendation reason;

[0020] Once the carrier proposal is approved by the user, it will proceed to the next stage of the process.

[0021] Specifically, the visual mapping table generation and gating module is configured as follows:

[0022] Each scientific mechanism in the knowledge file is numbered, and a corresponding visual indicator and carrier translation scheme are generated for each scientific mechanism, forming a mapping entry containing a triple of scientific mechanism number, visual indicator and carrier translation scheme, to generate a visual mapping table.

[0023] The visual mapping table number is reused as a check index in subsequent verification processes;

[0024] The version number of the visual mapping table is incorporated into the state data structure.

[0025] Specifically, in the meta-task state card persistence module:

[0026] The state data structure is a versioned data structure that includes target elements, constraint elements, evaluation criteria elements, and human-machine division of labor elements.

[0027] The target elements include carrier track, reader level, narrative mode and thinking model;

[0028] The constraint elements include the original text of the art style anchoring statement, the original text of the character locking statement, and the status of the setting table.

[0029] The assessment criteria elements include the visual mapping table version number and the checklist version;

[0030] The human-machine division of labor elements include the level of autonomy, collaboration stage markers, current page number, and previous page diagnostic summary.

[0031] The page-level cooperative loop engine forcibly reads the complete state data structure of the current version before generating each page;

[0032] Any changes to a state data structure field must be confirmed by the user and the version number must be incremented.

[0033] When resuming a session, a handshake protocol is executed, and production can only resume after each field of the status data structure is reviewed and confirmed.

[0034] Specifically, in the page-level cooperative loop engine, each iteration contains six phases executed sequentially:

[0035] Read the phase, forcibly reading the full text of the current version's state data structure;

[0036] The proposal phase generates a mind model layout, visual description, and checklist for the current page based on the knowledge points on the current page.

[0037] The gate phase is determined based on the current level of autonomy, deciding whether to pause and wait for manual confirmation or resume automatically.

[0038] Generate a phase, assemble a cue word package containing anchor fields, and call an image generation engine to generate an image, wherein the anchor fields are copied word-for-word from the constraint elements of the state data structure;

[0039] The phase is checked, and the checking and differential correction module is called to judge the generated image;

[0040] Telemetry and propulsion phases: Record the verification results for this page and pass the diagnostic summary for this page to the proposal phase on the next page.

[0041] Specifically, in the verification and differentiation correction module:

[0042] The multi-level classification includes a first category - clearly presented, a second category - partially presented, and a third category - missing or incorrect. The condition for passing this page is that all check items are judged to be in the first category.

[0043] The correction strategy includes: triggering a local correction that keeps the main structure of the page unchanged when a second category determination exists, and triggering a full-page reconstruction correction that rewrites the descriptive prompts into structured constraints when a third category determination exists;

[0044] If both a second-category determination and a third-category determination exist simultaneously, the full-page reconstruction correction triggered by the third-category determination will be executed first.

[0045] Furthermore, the preset termination condition is:

[0046] If the number of correction rounds on the same page reaches a preset threshold and still fails to pass, the upgrade ladder is triggered;

[0047] The upgrade ladder includes a three-level progressive strategy: the first level is to change the strategy, which means to stop repeated revisions of single pages and establish new standards applicable to the whole book; the second level is to change the tool, which means to switch the image generation engine; and the third level is to hand over the decision to human, which means to suspend the automatic process and hand it over to human decision-making.

[0048] When entering the second level of the upgrade ladder, the autonomy level is automatically downgraded.

[0049] Specifically, in the autonomy regulator:

[0050] The multiple levels of manual intervention density include page-by-page confirmation, batch confirmation, and exception interruption levels.

[0051] The objective metrics include the number of pages passed consecutively in one go and the upgrade ladder trigger events;

[0052] When N consecutive pages pass the initial review, a suggestion to upgrade to the next level is sent to the user, and the upgrade is implemented after the user's explicit consent.

[0053] When the second level of the upgrade ladder is triggered, the system automatically downgrades by one level and informs the user of the reason.

[0054] Gear changes are recorded in the status data structure as a human-machine division of labor element.

[0055] Preferably, the system further includes a VPT telemetry and reflection checkpoint module:

[0056] The VPT telemetry and reflection checkpoint module is used to immediately save the audit trail to disk at the end of each page's verification, including four columns: diagnostic findings, revision focus, screen changes, and manual review results. When the entire book is completed, it is summarized and a collaborative stage evolution summary is output.

[0057] The system outputs reflective questions to users at a preset page count and rhythm, pointing to the evaluation criteria themselves, and records user responses in the audit track.

[0058] When a user responds and confirms the need to revise the evaluation criteria, the version upgrade of the visual mapping table and the corresponding update of the evaluation criteria elements in the state data structure are triggered to achieve book-level reflow.

[0059] A second aspect of this invention provides a human-computer collaborative generation method for scientific text and image based on visual mapping verification, comprising the following steps:

[0060] Knowledge initiation and carrier determination: acquiring knowledge documents and determining the carrier scheme for graphic and textual publications;

[0061] Visual mapping table generation: The knowledge content in the knowledge file is converted into a visual mapping table.

[0062] Persistence of meta-task status cards organizes the fixed attributes and constraints of this text and image generation task into a status data structure for persistent storage.

[0063] A page-level iterative loop is executed sequentially for each page, including generation and verification. The generation is anchored by a state data structure, and the verification is based on a visual mapping table to determine the generated content.

[0064] Verification and correction: The judgment results are classified into multiple levels, and the correction strategy at the corresponding level is executed according to the classification results. The correction loop is terminated after the preset termination condition is triggered.

[0065] The system allows for autonomous adjustment, setting multiple levels of human intervention density and driving the shift of these levels based on objective indicators of loop operation.

[0066] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention uses visual mapping table hard gating to carry epistemological filtering of subject knowledge, uses meta-task state card to persist human-computer collaboration to define the weight, uses three-value verification and promotion / demotion rules to drive a terminateable and controlled generation-evaluation-correction loop, and uses VPT telemetry within the loop to ensure the process is auditable, thereby efficiently generating scientifically accurate, consistent across pages and supporting the development of teachers' meta-task awareness in a non-programming environment. (2) By setting up a visual mapping table generation and gating module, this invention establishes an objective, verifiable and unavoidable mandatory intermediary between scientific mechanisms and visual expression. The visual mapping table number is reused as a check index in the subsequent verification process and the version number is included in the status data structure. This formalizes the mapping relationship that is at the level of the paper table in academic research into a system hard gate, ensuring the accuracy and traceability of the visual expression of scientific content from the perspective of etiology. (3) By setting up a persistent module for the meta-task status card, the present invention designs the status data structure as a data structure with version number and containing target elements, constraint elements, evaluation criteria elements and human-machine division of labor elements. The page-level collaborative loop engine forces the reading of the complete status data structure of the current version before each page is generated. Any field change must be confirmed by the user and the version number must be incremented. When continuing across sessions, the handshake protocol is executed to repeat and confirm each field before production can be resumed. Thus, the persistent data structure replaces the easily decaying dialogue memory, fundamentally solving the cross-page consistency problem caused by the drift of role locking statements and style anchoring statements with the context distance in long sessions and cross sessions. (4) This invention sets up a multi-level classification verification and correction strategy, and uses the condition that all check items are clearly presented as the passing condition of this page. When there is partial presentation of the category, local correction is triggered. When there is missing or incorrect category, the whole page is reconstructed and corrected. This makes the correction intensity and deviation type accurately matched, avoiding the inefficiency and over-regeneration problems caused by the rough correction method of regenerating the whole page regardless of the error type in the traditional solution. (5) The present invention provides a step-by-step enhanced termination guarantee mechanism for the correction loop by setting a termination condition. When the number of correction rounds on the same page reaches a preset threshold and still fails to pass, it triggers a three-level progressive upgrade ladder including changing strategy, changing tools, and handing over to the decision-maker. This fundamentally solves the risk of infinite loop caused by the lack of termination conditions in the correction loop in the prior art. At the same time, it transforms the correction failure into a trigger condition for standard upgrade, realizing a qualitative change from passive retry to active optimization. (6) By setting an autonomy regulator, the present invention adopts three levels of manual intervention density: page-by-page confirmation, batch confirmation, and exception interruption. This achieves a controlled dynamic balance between efficiency, controllability, and user subjectivity. It avoids the inefficiency problem of the interaction cost increasing linearly with the number of pages caused by full manual confirmation. It also prevents the risk of loss of control and hidden dependence caused by complete automation through the dual protection of user authorization for advancement and automatic downgrading in case of failure. (7) By setting up a VPT telemetry and reflection checkpoint module, the present invention immediately saves the audit trajectory containing four columns of diagnostic findings, revision focus, screen changes and manual review results to the disk at the end of each page's verification. This changes the collaborative trajectory from post-event reconstruction to in-loop instant generation. The audit trajectory can be used as the original material for teacher training and process data for educational research, breaking through the technical limitations of existing technologies where collaborative trajectories are not auditable and cannot support the reuse of teaching research. At the same time, the module outputs reflection questions pointing to the assessment criteria itself to the user according to the page number rhythm and records the user's response into the audit trajectory. When the user's response confirms that the assessment criteria need to be revised, it triggers the version upgrade of the visual mapping table and the update of the assessment criteria elements in the status data structure to achieve book-level backflow. This allows the assessment results not only to be used for the correction of the current page, but also to flow back to the design layer to promote the iterative optimization of the assessment criteria itself. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0068] Figure 1This is an architecture diagram of the scientific image and text human-computer collaborative generation system based on visual mapping verification according to the present invention.

[0069] Figure 2 This is a schematic diagram of the data structure and lifecycle process of the meta-task status card of the present invention.

[0070] Figure 3 This is a six-phase flowchart of the page-level cooperative loop of the present invention.

[0071] Figure 4 This is a flowchart of the three-value verification decision tree of the present invention.

[0072] Figure 5 This is a schematic diagram of the VPT four-column audit trajectory and reflection checkpoint feedback of the present invention.

[0073] Figure 6 This is a schematic diagram of the phased autonomy progression state machine of the present invention.

[0074] Figure 7 This is a comparison diagram of visual mapping table entries and printed pages in an embodiment of the present invention.

[0075] Figure 8 This is a statistical chart of 16 sets of cross-page, page-by-page collaborative loop data in an embodiment of the present invention.

[0076] Figure 9 This is a schematic diagram of the evolution of evaluation criteria and the human-machine division of labor in an embodiment of the present invention. Detailed Implementation

[0077] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0078] Reference Figure 1 , Figure 1 The diagram illustrates the correspondence between the seven modules and the three layers (Foundation / Engine / Process) of the MTA-TPACK spiral. This invention instantiates the MTA-TPACK dynamic collaboration spiral from a research process engineering model that relies on manual expert intervention into a deployable system. Its overall architecture includes a static knowledge layer (corresponding to TPACK Foundation), a meta-task state layer and gating (corresponding to the persistence of the MTA navigation engine), a page-level collaboration loop (corresponding to the Generate-Evaluate-Refine spiral), and a telemetry layer (corresponding to VPT). The system runs on a conversational large model platform, and both the model and the image generation engine are pluggable. The image generation engine is adapted via an abstract interface, allowing for pluggable replacement.

[0079] The first aspect of this invention provides a scientific image and text human-computer collaborative generation system based on visual mapping verification, comprising:

[0080] Module 1: Knowledge Initiation and Carrier Determination Module, used to acquire knowledge documents and determine the carrier scheme for graphic and textual publications.

[0081] The knowledge initiation and carrier determination module is configured as follows:

[0082] The system prioritizes reading knowledge files provided by the user; if no knowledge files are available, it automatically retrieves and summarizes the core mechanisms and submits them for user confirmation.

[0083] Based on a four-dimensional judgment table of reader level, knowledge structure, emotional goals and material form, output one primary and one backup carrier proposal including story picture book track and science magazine illustrated book track, with each carrier proposal accompanied by a recommendation reason;

[0084] Once the carrier proposal is approved by the user, it will proceed to the next stage of the process.

[0085] This module avoids narrative frameworks that may lead to scientific expression bias by introducing subject knowledge as an epistemological filter for decision-making at the knowledge initiation stage, thus ensuring scientific accuracy from the image verification layer to the carrier decision-making layer.

[0086] Module 2: Visual Mapping Table Generation and Gating Module (Core Intermediary), which is used to convert the knowledge content in the knowledge file into a visual mapping table. The visual mapping table can only enter the subsequent image generation process after the user confirms each item.

[0087] The visual mapping table generation and gating module is configured as follows:

[0088] Each scientific mechanism in the knowledge file is numbered, and a corresponding visual indicator and carrier translation scheme are generated for each scientific mechanism, forming a mapping entry containing a triple of scientific mechanism number, visual indicator and carrier translation scheme, to generate a visual mapping table.

[0089] The visual mapping table number is reused as a check index in subsequent verification processes;

[0090] The version number of the visual mapping table is incorporated into the state data structure.

[0091] This module formalizes the visual indicator mapping into a hard gate of the system. Without the user's confirmation of each indicator, the system is prohibited from entering the storyboard and image generation process. The visual mapping table number is reused as a check index in the subsequent verification process, and the version number is included in the status data structure. This fundamentally prevents the problem of directly generating images by bypassing the mapping table, and establishes an objective, verifiable, and unavoidable mandatory intermediary between scientific mechanisms and visual expression.

[0092] Module 3: Meta-Task State Card Persistence Module, which is used to organize the fixed attributes and constraints of this image and text generation task into a state data structure for persistent storage.

[0093] The state data structure is a versioned data structure that includes target elements, constraint elements, evaluation criteria elements, and human-machine division of labor elements.

[0094] The target elements include carrier track, reader level, narrative mode and thinking model;

[0095] The constraint elements include the original text of the art style anchoring statement, the original text of the character locking statement, and the status of the setting table.

[0096] The assessment criteria elements include the visual mapping table version number and the checklist version;

[0097] The human-machine division of labor elements include the level of autonomy, collaboration stage markers, current page number, and previous page diagnostic summary.

[0098] The page-level cooperative loop engine forcibly reads the complete state data structure of the current version before generating each page;

[0099] Any changes to a state data structure field must be confirmed by the user and the version number must be incremented.

[0100] When resuming a session, a handshake protocol is executed, and production can only resume after each field of the status data structure is reviewed and confirmed.

[0101] Reference Figure 2 , Figure 2 This is a flowchart of the data structure (four-element partition) and lifecycle (first release → reprint → version increment → cross-session handshake) of the meta-task status card in this invention. This module replaces the easily decaying dialogue memory with a persistent data structure. The status card is organized according to the four elements of MTA theory. Any field change must be confirmed by the user and the version number must be incremented. The original text is forcibly reprinted every page. Cross-session continuation executes the handshake protocol and repeats the confirmation item by item before production can resume. This fundamentally solves the cross-page consistency problem caused by the drift of role locking statements and style anchoring statements with context distance in long sessions and cross sessions.

[0102] Module 4: Page-level Collaborative Loop Engine (Engineering of GER Spiral), which is used to execute an iterative loop containing generation and verification for each page. The generation is anchored to a state data structure, and the verification is based on a visual mapping table to determine the generated content.

[0103] In the page-level cooperative loop engine, each iteration contains six phases executed sequentially:

[0104] Read the phase, forcibly reading the full text of the current version's state data structure;

[0105] The proposal phase generates a mind model layout, visual description, and checklist for the current page based on the knowledge points on the current page.

[0106] The gate phase is determined based on the current level of autonomy, deciding whether to pause and wait for manual confirmation or resume automatically.

[0107] Generate a phase, assemble a cue word package containing anchor fields, and call an image generation engine to generate an image, wherein the anchor fields are copied word-for-word from the constraint elements of the state data structure;

[0108] The phase is checked, and the checking and differential correction module is called to judge the generated image;

[0109] Telemetry and propulsion phases: Record the verification results for this page and pass the diagnostic summary for this page to the proposal phase on the next page.

[0110] Reference Figure 3 , Figure 3 This is a six-phase flowchart of the page-level collaborative loop of the present invention, including gate branches, differentiated correction diversion, and upgrade ladder exits. This module uses a six-phase standardized loop to engineer the GER spiral in MTA-TPACK theory into a repeatable loop. The anchoring fields in the generated phase are copied word-for-word from the state card rather than relying on model memory. The telemetry and advancement phases pass the diagnostic summary of this page to the proposal phase of the next page, realizing the CoLDE chain design. This allows the evaluation experience to be accumulated and passed from page to page throughout the book, ensuring the coherence and consistency of the cross-page narrative.

[0111] Module 5: Tri-value verification and differential correction module, used to classify the judgment results into multiple levels, execute the corresponding level correction strategy according to the classification results, and terminate the correction loop after triggering the preset termination condition.

[0112] The multi-level classification includes a first category - clearly presented (√), a second category - partially presented (△), and a third category - missing or incorrect (×). All check items must be judged as the first category to pass this page.

[0113] The correction strategy includes: triggering a local correction that keeps the main structure of the page unchanged when a second category determination exists, and triggering a full-page reconstruction correction that rewrites the descriptive prompts into structured constraints when a third category determination exists;

[0114] If both a second-category determination and a third-category determination exist simultaneously, the full-page reconstruction correction triggered by the third-category determination will be executed first.

[0115] This module adopts and strengthens the √ / △ / × indicator standard of MTA-TPACK research in the scientific accuracy dimension: It extracts items from the mapping table and expands them into a checklist set C = {c1,…,ck}, where k is the number of checklist items. Each item is judged as ci ∈ {√ clearly presented, △ partially presented, × missing or incorrect}, with the condition ∀ci = √. △ items trigger local reinforcement instructions (position anchoring, contrast enhancement, local editing), while × items trigger structural reconstruction instructions (rewriting descriptive terms into structural constraints—arrow direction, layer order, label correspondence—and then regenerating). This ensures precise matching between correction intensity and deviation type, avoiding the inefficiency and excessive regeneration problems caused by the traditional approach of regenerating the entire page regardless of error type. Visual clarity, narrative integration, and reader age suitability are assessed using a 1–5 scale (threshold > 3).

[0116] The preset termination condition is:

[0117] If the number of correction rounds on the same page reaches a preset threshold and still fails to pass, the upgrade ladder is triggered;

[0118] The upgrade ladder includes a three-level progressive strategy: the first level is to change the strategy, which means to stop repeated revisions of single pages and establish new standards applicable to the whole book; the second level is to change the tool, which means to switch the image generation engine; and the third level is to hand over the decision to human, which means to suspend the automatic process and hand it over to human decision-making.

[0119] When entering the second level of the upgrade ladder, the autonomy level is automatically downgraded.

[0120] Reference Figure 4 , Figure 4 This invention employs a three-valued verification decision tree (√ / △ / × → pass / local reinforcement / structural reconstruction → upgrade ladder). It sets a correction round counter r. If r≥2 on the same page and all pages are still not √, the three-level upgrade ladder is triggered (change strategy → change tool → hand over to human decision-maker). This ensures the loop will terminate. After a correction failure, the same strategy is not repeated; instead, it upgrades to a book-level specification, benefiting all subsequent pages. This fundamentally solves the risk of infinite loops caused by the lack of termination conditions in existing technologies, while transforming correction failures into opportunities for specification upgrades, achieving a qualitative leap from passive retries to proactive optimization.

[0121] Module Six: VPT Telemetry and Reflective Checkpoint Module, used to immediately save the audit trail to disk at the end of each page's verification. The trail includes four columns: diagnostic findings, revision focus, screen changes, and manual review results. When the entire book is completed, the track is summarized and a collaborative stage evolution summary is output.

[0122] This module outputs reflective questions to the user at a preset page count, pointing to the evaluation criteria themselves, and records the user's responses in the audit track.

[0123] When a user responds and confirms the need to revise the evaluation criteria, the version upgrade of the visual mapping table and the corresponding update of the evaluation criteria elements in the state data structure are triggered to achieve book-level reflow.

[0124] Reference Figure 5 The format of the VPT four-column audit trail (diagnostic findings | revision focus | screen changes | human review results) is similar to the cyclical audit table in the MTA-TPACK study (Appendix D of the original paper: "diagnosis → prompt revision → observed change → human review"). The system is isomorphic to "verification", but changes from "researcher post-event compilation" to "instant on-the-fly within the loop". When the book is completed, it only summarizes and outputs a summary of the evolution of the collaboration stage. In the research log mode, it synchronously outputs structured JSON (including revision_code, autonomy_level, and collab_stage fields), so that the collaboration trajectory is changed from post-event reconstruction to instant generation within the loop, eliminating the selection bias and information distortion caused by post-event compilation. The reflection checkpoint raises reflection questions to users on the evaluation criteria itself according to the page number rhythm (every N pages completed, 8 pages by default) (such as "Does the most △ indicator indicate that the mapping table criteria need to be revised?"). The user response is recorded in VPT. When the evaluation criteria revision is triggered, the book-level backflow gate is activated, so that the evaluation results flow back to the design layer to drive the iteration and optimization of the criteria. At the same time, by forcing users to respond to reflection questions on the criteria itself, it effectively combats the implicit user dependence that the system scaffolding may cause.

[0125] Module 7: Staged Autonomy Regulator, used to set three levels of artificial intervention density. This corresponds to the assessment of labor redistribution spectrum in the MTA-TPACK study, which is "human-led collaboration → AI closed-loop self-assessment". The system also marks the current collaboration stage and records it in the status card according to the four stages of AI self-assessment maturity in the study (descriptive comparison → diagnostic assessment → structured introspection → selective stability). The system drives gear shifting based on objective indicators of loop operation, and gear advancement requires user authorization.

[0126] The multiple levels of manual intervention density include page-by-page confirmation, batch confirmation, and exception interruption levels.

[0127] The objective metrics include the number of pages passed consecutively in one go and the upgrade ladder trigger events;

[0128] When N consecutive pages pass the initial review, a suggestion to upgrade to the next level is sent to the user, and the upgrade is implemented after the user's explicit consent.

[0129] When the second level of the upgrade ladder is triggered, the system automatically downgrades by one level and informs the user of the reason.

[0130] Gear changes are recorded in the status data structure as a human-machine division of labor element.

[0131] Reference Figure 6 , Figure 6 This invention provides a phased autonomy progression / degradation state machine ( (Migration conditions and manual consent nodes). This module achieves a controlled dynamic balance between efficiency, controllability, and user autonomy. Promotion is based on the number of consecutive pages passed (e.g., passing 4 consecutive pages at once) and requires explicit user consent to prevent the system from expanding its authority on its own. Demotion is triggered by entering the second level of the upgrade ladder and is automatically executed by the system, with the reason informed to the user. The change in rank is recorded in the status card and VPT, forming an observable research variable. This avoids the inefficiency of full manual confirmation and prevents the risk of loss of control and hidden dependence caused by complete automation through the dual safeguards of user authorization for promotion and automatic demotion in case of failure.

[0132] A second aspect of this invention provides a human-computer collaborative generation method for scientific text and image based on visual mapping verification, comprising the following steps:

[0133] Knowledge initiation and carrier determination: acquiring knowledge documents and determining the carrier scheme for graphic and textual publications;

[0134] Visual mapping table generation: The knowledge content in the knowledge file is converted into a visual mapping table.

[0135] Persistence of meta-task status cards organizes the fixed attributes and constraints of this text and image generation task into a status data structure for persistent storage.

[0136] A page-level iterative loop is executed sequentially for each page, including generation and verification. The generation is anchored by a state data structure, and the verification is based on a visual mapping table to determine the generated content.

[0137] Verification and correction: The judgment results are classified into multiple levels, and the correction strategy at the corresponding level is executed according to the classification results. The correction loop is terminated after the preset termination condition is triggered.

[0138] The degree of autonomy is adjusted, with multiple levels of human intervention density set, and the level shift is driven based on objective indicators of loop operation.

[0139] VPT telemetry and reflection checkpoints are immediately logged at the end of each page's verification, containing a four-column audit trail including diagnostic findings, revision focus, screen changes, and manual review results. When the entire book is completed, a summary of the collaborative stage evolution is output.

[0140] The specific implementation methods of the present invention will be described in detail below with reference to specific implementation examples.

[0141] Example 1: A science book about tornadoes, "The Last Mile of a Tornado (Part 1)"

[0142] Implementation Overview: The knowledge document is a review text titled "Tornado Anatomy: Twenty Years of Tornado Science," co-authored by the authors and a large language model. It covers nearly twenty years of research on tornado formation theory, observation, simulation forecasting, and risk management. The implementation platform is a dialogic large model, with a raw image engine adapting to seven-field prompts for word pack matching. After image generation and return, a three-value check is performed. The output is the first volume of the four-volume standard edition series—"The Last Mile of Tornadoes (Part 1): Spinning, But Not Enough," comprising 32 pages plus a cover, 3:4 vertical format, for readers at level L4 (upper elementary school, approximately 10-12 years old). The medium is a hybrid science magazine / picture book format, with the narrative mode primarily using N17 cross-sectional diagrams and secondarily using N18 question-and-answer sessions. The main thinking model is a vertical causal construction. The style anchoring statement (original text, word-for-word entered into the status card) is: "Natural history realistic scientific illustration style, composed of dark gray-blue storm clouds, warm orange flowing water, cyan-blue cold pools, and white streamlines; the three-dimensional cross-section has a realistic atmospheric scale and soft volumetric light; all arrows, layers, labels, and height scales are clear and consistent; scientific structure takes precedence over dramatic effect; 3:4 vertical scientific illustration book layout." In this embodiment, the above style anchoring statement is stored word-for-word into the status card and is forcibly loaded during the generation of each subsequent page, ensuring the consistency of the style across pages from the perspective of the cause of the problem.

[0143] Knowledge Initiation and Carrier Determination (Module 1 Transcript): After performing knowledge structure determination on the review text, the system first outputs three volume options (three-volume concise version / four-volume standard version / five-volume research version) with applicable scenarios. After user confirmation, the four-volume standard version was adopted. The core question of the first volume is defined as "Why do many supercells rotate, but only a few eventually form tornadoes?" The carrier determination outputs the main recommendation (a hybrid track of science magazine and picture book, N17+N18) and alternatives (guided adventure track, N15+N10), each with reasons; and proactively provides counterexamples for exclusion: "It is not recommended to make tornadoes the anthropomorphic protagonists who can think and actively choose paths, because it is easy to depict the controversial SVC causal relationship as a definite intention." The subject knowledge, as an epistemological filter, has already taken effect at the carrier decision-making level, rather than just at the image verification level.

[0144] Visual mapping table hard gating (Module 2 record): The system produces a mapping table VM-TOR-01 v1.0, with a total of 10 entries (T1-T10), each being a quadruple of (scientific mechanism, observable visual indicator, picture book track translation, and magazine / illustrated track translation). Representative entries: T3 "Mid-level cyclones are usually more noticeable several kilometers above the ground, but this does not mean that a tornado has already formed" → visual indicator "broad rotation in the cloud; no narrow rotating column on the ground yet, and a clear distance is maintained between the two", magazine track translation with a prohibitive constraint "the funnel is prohibited from directly connecting to the ground"; T6 vorticity stretching entry allows the analogy of "skaters tightening their arms" in picture book track translation but mandates the annotation "this is only an angular momentum analogy"; T8 (SVC controversy mechanism) stipulates "use dashed lines or question marks to indicate uncertain causality, and prohibits the claim that a tornado will inevitably form with a single solid arrow" - formalizing scientific uncertainty itself into a verifiable visual indicator. After each entry in the mapping table is confirmed by the user, its version number is recorded in the status card (hard-gated). T1-T6 are the main check indexes for the first volume, and T7 serves as a final announcement. (See reference...) Figure 7 , Figure 7 The correspondence between entry T3 and the title page, cover, and the enlarged cyclone pages on pages 15-16 is shown.

[0145] Meta-task Status Card (Module 3 Record): The status card PSC-TOR-B1U was first released as v2.0 during the planning and finalization process. Fields are organized according to four elements: "Objective," "Constraints," "Evaluation Criteria," and "Human-Machine Division of Labor." The objective element includes the book title and carrier track, L4 level, N17+N18 narrative mode, and vertical causal construction thinking model. The constraint element includes the style anchoring to the original text, the original text of each role's locked statements, and the status of the setting table. The evaluation criteria element includes the VM-TOR-01 v1.0 version number and confirmation date, and the three-value checklist TC-TOR-B1U v1.0. The human-machine division of labor element includes autonomy A0, the "description comparison" of the collaboration stage, the current page number, and the diagnostic summary of the previous page. The status card is reprinted verbatim for the first step of each cross-page loop. The entire version evolution is from v2.0 to v2.7, with each change (such as volume structure adjustment, mapping table upgrade, and current page and round number update) being incremented after user confirmation.

[0146] Page-level Collaborative Loops and Differentiation Correction (Modules Four and Five): All 16 cross-page sets in the book execute a six-phase loop: ① Read (reprint status card) → ② Proposal (knowledge points, mind model layout, visual description, and three-value check items for this page) → ③ Gate (based on autonomy level) → ④ Generate (seven-field prompt word package, anchor fields copied word-for-word from status card) → ⑤ Verify → ⑥ Telemetry and Advancement. Proposals on subsequent pages must reference the diagnostic findings of this page, realizing the chain design of CoLDE.

[0147] Three representative cases:

[0148] Case A (passed on the first try): Pages P3-P4 “Two storms, looking almost identical”, T7 three-value check all √, scale 4-5 points, correction round r=0, VPT drop plate chain advance.

[0149] Case B (△→Local Enhancement): On pages 15-16, the enlarged cyclone page was checked. 7√ and 1△ were checked ("The enlarged image still shows a funnel-shaped outline that is wider at the top and narrower at the bottom, which can easily be mistaken for a giant tornado"). No × were found. Local enhancement was adopted according to the differentiated correction rules (widening the lower outline, weakening the near-ground convergence, and anchoring the label to the 3-10 km altitude zone). The entire page was not regenerated. The case passed and a △ item was attached.

[0150] Case C (×→Structural Restructuring→Upgrade Ladder): The first round of checks on the P1 title page is 1× ("the page reveals subsequent mechanisms in advance"), triggering structural restructuring (deleting all mechanism labels and explanation boxes, weakening the central vertical cloud column, and retaining only one question), r=1 passed. The first round of checks on the P9-P10 wind shear pages is 1×+1△ (the lower-level "northward horizontal wind" is drawn as a vertical upward arrow, which is easily misread as an updraft) → after r=1 structural restructuring, it is still 7√2△ → after r=2 partial editing, it is still 1△ (the same arrow shape carries two scientific meanings in the whole book - it is a graphic language problem rather than a knowledge error) → according to the rule of "two rounds of checks on the same page are not complete", it officially enters the upgrade ladder Level 1 (change strategy): the page is no longer reborn, but a book-level "dual coordinate graphic standard" is established (the arrow in the cross-section only represents the vertical direction, the geographical location is always expressed by the top view / compass, and the same arrow shape cannot carry two scientific meanings), P9-P10 is judged to pass accordingly. This event also constituted an evaluation standard revision: the visual mapping table was upgraded to VM-TOR-01v1.1, and a new entry G1 "Graphics Language Specification" was added, which benefited all subsequent pages. The actual test verified the two design goals of "inevitable termination" and "standard reflow" of the upgrade ladder.

[0151] Evolution of Assessment Standards and Reflection Checkpoints (Module Six Transcript): In addition to G1, the collaboration also summarized the book-level standard G2 "One Split Page = One Scientific Action" (established because the most common △ item is "Images prematurely reveal the answer on the next page"). After all the inner pages of the book are completed but before the cover is generated, the system triggers the teacher reflection checkpoint, posing a question to the user that points to the assessment standards themselves: "Will the second volume continue to strictly adhere to 'revealing only one mechanism per split page, even if it reduces the drama of a single page'?" After the user confirms, it is recorded in VPT, and G2 is solidified as a series-level standard. All four columns of VPT entries for each page of the book (diagnostic findings / revision focus / image changes / human review results) are immediately logged in phase 6 of each page loop, and only a summary is made at the end of the project.

[0152] Actual test results data: (Refer to) Figure 8 and Figure 9 , Figure 8The data for 16 sets of cross-page, page-by-page collaborative loops are shown (distribution of the number of correction rounds, first pass, and correction type for each cross-page). Figure 9 The evolution of evaluation criteria and the trajectory of human-machine division of labor are shown (mapping table v1.0→v1.1, G1 / G2 book-level specifications, reflection checkpoints, collaborative stage advancement, and full autonomy A0).

[0153] The actual test results are shown in Table 1 below:

[0154] Table 1. Measured Results Data

[0155] The conformity comparison between the actual measurement process and the technical solution is shown in Table 2 below:

[0156] Table 2. Comparison of Actual Measurement Process and Technical Solution

[0157] The above-mentioned test data fully verified the effectiveness of each module of the present invention in actual deployment. The visual mapping table hard gating intercepted scientific illusions at the source; the status card persistence maintained the consistency of the style and characters throughout the 16 pages of the book; the three-value verification and differential correction enabled 81% of the pages to pass on the first try, with the correction intensity and deviation type accurately matched; the upgrade ladder terminated the infinite correction loop after being triggered in P9-P10 and generated a book-level specification that benefits all subsequent pages; the advancement guardrail of the autonomy regulator strictly took effect as designed; and the instantaneous disk drop within the VPT trajectory loop eliminated post-reconstruction distortion.

[0158] The terms used in this invention are explained in Table 3 below:

[0159] Table 3. Terminology Explanation

[0160] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A scientific image and text human-computer collaborative generation system based on visual mapping verification, characterized in that, include: The knowledge initiation and carrier determination module is used to acquire knowledge files and determine the carrier scheme for graphic and textual publications; The visual mapping table generation and gating module is used to convert the knowledge content in the knowledge file into a visual mapping table. The visual mapping table can only enter the subsequent image generation process after the user confirms each item in the visual mapping table. The meta-task status card persistence module is used to organize the fixed attributes and constraints of this image and text generation task into a status data structure for persistent storage. The page-level collaborative loop engine is used to execute an iterative loop containing generation and verification for each page. The generation is anchored by a state data structure, and the verification is based on a visual mapping table to determine the generated content. The verification and differentiation correction module is used to classify the judgment results into multiple levels, execute the corresponding level correction strategy according to the classification results, and terminate the correction loop after triggering the preset termination condition. The autonomy regulator is used to set multiple levels of human intervention density. It drives the shift of levels based on objective indicators of loop operation, and the upgrade requires user authorization.

2. The scientific image and text human-computer collaborative generation system based on visual mapping verification as described in claim 1, characterized in that, The knowledge initiation and carrier determination module is configured as follows: The system prioritizes reading knowledge files provided by the user; if no knowledge files are available, it automatically retrieves and summarizes the core mechanisms and submits them for user confirmation. Based on a four-dimensional judgment table of reader level, knowledge structure, emotional goals and material form, output one primary and one backup carrier proposal including story picture book track and science magazine illustrated book track, with each carrier proposal accompanied by a recommendation reason; Once the carrier proposal is approved by the user, it will proceed to the next stage of the process.

3. The scientific image and text human-computer collaborative generation system based on visual mapping verification as described in claim 1, characterized in that, The visual mapping table generation and gating module is configured as follows: Each scientific mechanism in the knowledge file is numbered, and a corresponding visual indicator and carrier translation scheme are generated for each scientific mechanism, forming a mapping entry containing a triple of scientific mechanism number, visual indicator and carrier translation scheme, to generate a visual mapping table. The visual mapping table number is reused as a check index in subsequent verification processes; The version number of the visual mapping table is incorporated into the state data structure.

4. The scientific image and text human-computer collaborative generation system based on visual mapping verification as described in claim 3, characterized in that, In the meta-task status card persistence module: The state data structure is a versioned data structure that includes target elements, constraint elements, evaluation criteria elements, and human-machine division of labor elements. The target elements include carrier track, reader level, narrative mode and thinking model; The constraint elements include the original text of the art style anchoring statement, the original text of the character locking statement, and the status of the setting table. The assessment criteria elements include the visual mapping table version number and the checklist version; The human-machine division of labor elements include the level of autonomy, collaboration stage markers, current page number, and previous page diagnostic summary. The page-level cooperative loop engine forcibly reads the complete state data structure of the current version before generating each page; Any changes to a state data structure field must be confirmed by the user and the version number must be incremented. When resuming a session, a handshake protocol is executed, and production can only resume after each field of the status data structure is reviewed and confirmed.

5. The scientific image and text human-computer collaborative generation system based on visual mapping verification as described in claim 4, characterized in that, In the page-level cooperative loop engine, each iteration contains six phases executed sequentially: Read the phase, forcibly reading the full text of the current version's state data structure; The proposal phase generates a mind model layout, visual description, and checklist for the current page based on the knowledge points on the current page. The gate phase is determined based on the current level of autonomy, deciding whether to pause and wait for manual confirmation or resume automatically. Generate a phase, assemble a cue word package containing anchor fields, and call an image generation engine to generate an image, wherein the anchor fields are copied word-for-word from the constraint elements of the state data structure; The phase is checked, and the checking and differential correction module is called to judge the generated image; Telemetry and propulsion phases: Record the verification results for this page and pass the diagnostic summary for this page to the proposal phase on the next page.

6. The scientific image and text human-computer collaborative generation system based on visual mapping verification as described in claim 1, characterized in that, In the verification and difference correction module: The multi-level classification includes a first category - clearly presented, a second category - partially presented, and a third category - missing or incorrect. The condition for passing this page is that all check items are judged to be in the first category. The correction strategy includes: triggering a local correction that keeps the main structure of the page unchanged when a second category determination exists, and triggering a full-page reconstruction correction that rewrites the descriptive prompts into structured constraints when a third category determination exists; If both a second-category determination and a third-category determination exist simultaneously, the full-page reconstruction correction triggered by the third-category determination will be executed first.

7. The scientific image and text human-computer collaborative generation system based on visual mapping verification as described in claim 6, characterized in that, The preset termination condition is: If the number of correction rounds on the same page reaches a preset threshold and still fails to pass, the upgrade ladder is triggered; The upgrade ladder includes a three-level progressive strategy: the first level is a strategy change, which means terminating repeated revisions on single pages and establishing new standards applicable to the entire book; the second level is a tool change, which means switching the image generation engine. The third level is to hand over the decision to a human, which means to suspend the automatic process and hand it over to a human. When entering the second level of the upgrade ladder, the autonomy level is automatically downgraded.

8. The scientific image and text human-computer collaborative generation system based on visual mapping verification as described in claim 7, characterized in that, In the autonomy regulator: The multiple levels of manual intervention density include page-by-page confirmation, batch confirmation, and exception interruption levels. The objective metrics include the number of pages passed consecutively in one go and the upgrade ladder trigger events; When N consecutive pages pass the initial review, a suggestion to upgrade to the next level is sent to the user, and the upgrade is implemented after the user's explicit consent. When the second level of the upgrade ladder is triggered, the system automatically downgrades by one level and informs the user of the reason. Gear changes are recorded in the status data structure as a human-machine division of labor element.

9. The scientific image and text human-computer collaborative generation system based on visual mapping verification as described in claim 1, characterized in that, The system also includes a VPT telemetry and reflection checkpoint module: The VPT telemetry and reflection checkpoint module is used to immediately save the audit trail to disk at the end of each page's verification, including four columns: diagnostic findings, revision focus, screen changes, and manual review results. When the entire book is completed, it is summarized and a collaborative stage evolution summary is output. The system outputs reflective questions to users at a preset page count and rhythm, pointing to the evaluation criteria themselves, and records user responses in the audit track. When a user responds and confirms the need to revise the evaluation criteria, the version upgrade of the visual mapping table and the corresponding update of the evaluation criteria elements in the state data structure are triggered to achieve book-level reflow.

10. A method for human-computer collaborative generation of scientific text and images based on visual mapping verification, implemented based on the human-computer collaborative generation system for scientific text and images according to any one of claims 1-9, characterized in that, Includes the following steps: Knowledge initiation and carrier determination: acquiring knowledge documents and determining the carrier scheme for graphic and textual publications; Visual mapping table generation: The knowledge content in the knowledge file is converted into a visual mapping table. Persistence of meta-task status cards organizes the fixed attributes and constraints of this text and image generation task into a status data structure for persistent storage. A page-level iterative loop is executed sequentially for each page, including generation and verification. The generation is anchored by a state data structure, and the verification is based on a visual mapping table to determine the generated content. Verification and correction: The judgment results are classified into multiple levels, and the correction strategy at the corresponding level is executed according to the classification results. The correction loop is terminated after the preset termination condition is triggered. The system allows for autonomous adjustment, setting multiple levels of human intervention density and driving the shift of these levels based on objective indicators of loop operation.