Teaching and evaluation integrated verifiable generation and causal optimization education intelligent system and method

Abstract

The invention discloses a verifiable generation and causal optimization education intelligent system and method oriented to teaching and evaluation integration. The system comprises a data access and learning portrait, a knowledge retrieval and knowledge graph, VGE (Verifiable Generation Environment), consistency / contradictory rate / coverage rate verification, multi-expert agent collaboration and arbitration, causal structure learning and anti-factual optimization, a context protocol and an edge-cloud double-loop scale block. The illusion is reduced through evidence chain binding and thresholding, and the content quality is guaranteed through multi-Agent voting and negative right; optimal intervention is selected based on a causal objective function, and a teaching plan, a calendar and a task list are written back in a closed loop, so that end-to-end optimization of question setting, reading, lecture and teaching intervention is realized, learning gain is improved, and fairness and teacher workload are considered.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of educational artificial intelligence technology, in particular to a verifiable generation and causal optimization education intelligent system and method for teaching and evaluation integration (classroom teaching-proposition-evaluation-feedback-lesson plan linkage), specifically comprising: evidence chain binding type generation (VGE) based on knowledge retrieval and knowledge graph, generation result consistency and contradiction rate checking, multi-expert agent collaboration and arbitration, learning portrait driven causal structure learning and counterfactual optimization, cross-link context protocol (MCP) and edge-cloud collaborative compliance audit mechanism. BACKGROUND

[0002] Existing education AI systems are mostly focused on single-link functions such as automatic question generation, automatic grading, teaching document generation or classroom behavior analysis. Such solutions generally have the following problems: (1) lack of verifiability: large language models may have hallucinations, lack traceable evidence chains and consistency constraints; (2) cross-link fragmentation: context breaks between classroom state-proposition-evaluation-lesson plan-calendar-intervention strategies, making it difficult to achieve closed-loop optimization; (3) lack of causal effectiveness: teaching interventions and learning outcomes are mostly dependent on correlation, lacking causal structure and counterfactual assessment; (4) weak trusted governance: lacking multi-agent review and arbitration mechanisms, lacking edge-cloud collaborative privacy audit and version rollback mechanisms.

[0003] In summary, there is an urgent need for a system and method that can generate evidence chains, strictly check consistency, arbitrate with multiple agents, and optimize causally within the closed loop of teaching and evaluation integration. SUMMARY

[0004] (I) Invention purpose The present application aims to address the pain points of existing technology in the context of "classroom teaching-proposition-evaluation-feedback-lesson plan and teaching plan linkage" teaching and evaluation integration, and proposes a verifiable generation (VGE, Verifiable Generative Engine) and causal optimization collaborative driving education intelligent system and method to solve the following problems: (1) Lack of generation credibility Existing large language models (LLM) are prone to hallucinations and inconsistencies in question generation, scoring rules and lecture note generation, lacking evidence chain binding, consistency / contradiction rate quantification and thresholding mechanisms, making it difficult to review and audit the content.

[0005] (2) Cross-link fragmentation and context loss Classroom data, question banks, knowledge points, evaluation reports, teaching documents and teaching plans are often distributed in different systems and conversations, lacking context protocols that can be understood and tracked by machines, and cannot form a closed-loop optimization, making it difficult to implement version rollback and compliance audit.

[0006] (3) Intervention decisions lack causal basis Most teaching interventions are based on correlation or rules of thumb, lacking modeling and counterfactual evaluation of the causal structure between intervention and learning outcomes, making it difficult to select intervention programs that are optimal for the target group, as well as for overall fairness and cost.

[0007] (4) Lack of credible arbitration in collaborative decision-making Opinions from various parties, such as question setting, difficulty, grading, compliance, and teaching methods, often conflict. The lack of a multi-expert intelligent agent scoring fusion, conflict detection, and arbitration backoff mechanism makes it difficult to form an explainable and implementable unified decision.

[0008] Therefore, the overall objective of this invention is: (1) Construct a content production and gatekeeping mechanism with verifiable generation as the core based on evidence chain binding, quantify consistency, coverage and contradiction rate and set thresholds; (2) Construct a trusted collaboration mechanism with multi-expert intelligent agent collaboration, arbitration, and rollback as its core; (3) Construct an intervention selection mechanism with causal structure learning-counterfactual optimization as the core, taking into account learning gains, fairness, teacher workload and time constraints; (4) Construct an engineering support system with context protocols (MCP type) and edge-cloud dual-ring compliance as the core to ensure cross-link context consistency, traceability, auditability and privacy compliance.

[0009] (II) Technical Solution 1 System Solution A verifiable generative and causal optimization-oriented intelligent educational system for integrated teaching and assessment includes the following modules and the data / control flow between them: A. Data Access and Learning Profile Module Access classroom-side data (attention / interaction frequency / homework submission / quiz scores, etc.), academic affairs / course outlines, question banks / textbooks, historical assessment reports and learning trajectories; Build learner profiles (behavior, knowledge mastery, learning pace, error patterns), supporting the coexistence of vectorized features and structured features; Provides minimum necessary fields and de-identification strategies to ensure privacy compliance.

[0010] B. Knowledge Retrieval and Knowledge Graph Module Establish a knowledge graph that includes course objectives, knowledge points, question items, scoring rules, and reference materials (textbooks, supplementary materials, authoritative literature); Establish a vector retrieval index (such as dense vector + BM25 mixed sorting) to support Top-k recall and evidence fragment localization; Supports graph alignment: Maps key concepts in the question stem / answer / scoring criteria to knowledge point nodes.

[0011] C. Verifiable Generation (VGE) Module Under the constraints of knowledge retrieval results and graphs, questions, standard answers, scoring rules, handouts, classroom feedback texts, etc. are generated. Forced Evidence Chain (EC): Records fields such as doc_id, span, similarity sim, timestamp, version number, source hash, and contributing agent; Supports controlled prompts (difficulty target μ_d, knowledge point coverage target, format constraints, terminology consistency, formula checking, etc.).

[0012] D. Consistency and Conflict Rate Verification Module Calculate the consistency score s_cons([0,1]), the contradiction rate r_contra([0,1]), and the evidence coverage rate r_cover([0,1]). Typical thresholds: θ_c∈[0.70,0.90], θ_r∈[0.00,0.15], r_cover_min∈[0.60,0.85]; If the threshold is not reached, a rollback and rewrite or a second search will be triggered; comprehensive weighting of multi-dimensional consistency sub-indicators such as numerical values, units, boundary conditions, knowledge point mapping, and logical references is supported.

[0013] E. Multi-expert intelligent agent collaboration and arbitration module This includes: Question-setting Agent, Difficulty Assessment Agent, Grading Agent, Compliance Agent (including detection of academic misconduct / copyright / sensitive content), Teaching Methodology Agent (teaching methodology suitability / interpretive assessment), and Arbitration Agent; The candidate generators are scored and explained, using weighted fusion / voting and conflict detection. Example of constraint rules: Compliant agents have veto power; a low score due to a combination of question setting and difficulty triggers a "rewrite with reduced difficulty"; high scores in teaching methodology take precedence over aesthetic writing style; When the score variance > σ0 or the key agent is below the threshold, the arbitration fallback link is entered, and the rewriting agent performs targeted rewriting based on the diagnostic prompts and sends it back for re-evaluation.

[0014] F. Causal Structure Learning and Counterfactual Optimization Module Learn the causal graph G=(V,E) between sets of variables V (such as ATTN, HW_RATE, READING, QUIZ_FREQ, DIFFICULTY, SCORE, FAIRNESS); Constraint / scoring search based on historical data and prior teaching structure (such as DAG learning, NOTEARS class, causal discovery + knowledge prior fusion). Perform counterfactual evaluation on candidate intervention u (adjusting the number of questions / question types / pacing / lecture handout structure / differentiated assignments, etc.) y_cf = counterfactual(G, do(u)); Objective function: Obj = ΔS − λ1·Fair − λ2·Work − λ3·Time, where λ1, λ2, λ3 ∈ [0,1] are set according to the strategy; Output the optimal intervention and its expected gains and costs, and record them in the context and audit logs.

[0015] G. Context Protocol Module (MCP Class) Define a parsable context schema across all stages (course metadata, knowledge point mapping, scoring rubric ID, profile summary, evidence chain, version watermark, privacy tag); Maintain context persistence and replayability; support cross-model / cross-session / cross-task migration, and ensure semantic and audit consistency.

[0016] H. Edge-Cloud Dual-Ring Collaboration and Compliance Module Edge side: Low-latency identification and PII desensitization, only uploading anonymous features; Cloud-based: Perform heavy-duty retrieval, VGE generation, consistency calculation, causal inference, version auditing, and watermarking. Audit: Each generation / arbitration / intervention selection generates a non-repudiable audit entry {ctx_id, agent_trace, EC, metrics, decision, version_hash, signer}; It supports withdrawal and minimum available dataset strategies to meet privacy regulations and institutional audit requirements.

[0017] Key features of the system: VGE+ consistency thresholding, multi-agent arbitration rollback, causal-counterfactual optimization, contextual protocol and dual-loop compliance, constitute a combination of technical mechanisms that can be distinguished from traditional "functional splicing" educational AI solutions.

[0018] 2. Methodology and Scheme A verifiable generative and causal optimization educational intelligence method for integrated teaching and assessment includes the following steps: S1. Data Acquisition and Modeling Collect data from multiple sources, including classroom sessions, assignments, and assessments. After edge-side anonymization, integrate the data with academic affairs, syllabus, question banks, and textbooks to construct learner profiles and knowledge graphs.

[0019] S2. Retrieval and Context Construction Perform a Top-k retrieval in the knowledge base / graph based on the task query q, construct the context ctx = {document fragments, graph nodes, similarity, version}, and write it into the context protocol object CTX.

[0020] S3. Verifiable Generation of Evidence Chain Binding (VGE) Generate candidate content (title, answer, scoring rules, lecture notes, feedback) under the constraint of CTX, and at the same time generate the evidence chain EC, and record the version and watermark.

[0021] S4. Consistency / Contradiction Rate / Coverage Calculation and Thresholding Calculate s_cons, r_contra, r_cover. If s_cons < θ_c or r_contra > θ_r or r_cover < r_cover_min, then perform a fallback: secondary retrieval → controlled rewriting → re-evaluation until it passes or reports an error.

[0022] S5. Multi-expert Agent Scoring Fusion and Arbitration Fallback The agents for question creation, difficulty level, grading, compliance, teaching method, etc. score and explain the candidate content; if there is a conflict or the total score S < tau, enter the arbitration fallback chain, and the rewriting agent will perform a directed correction and re-evaluation.

[0023] S6. Causal Structure Learning and Counterfactual Optimization Learn or update the causal graph G on the rolling data of the portrait / log; perform counterfactual evaluations on several intervention candidates U = {u_i} and select the optimal intervention u* with the objective function Obj; output the intervention script and the expected effect.

[0024] S7. Context and Document Linkage Write the content passed through arbitration and the intervention plan back to CTX and synchronize them to the teaching plan, teaching calendar, learning task sheet, and evaluation report; update the version, generate the watermark, and audit evidence.

[0025] S8. Deployment and Closed-loop Rewriting Execute edge-cloud collaborative deployment; periodically collect the implementation effects and write them back to the portrait and G, triggering the next round of rolling optimization.

[0026] 3 Key Criteria, Parameters, and Data Structures (1)Consistency and Contradiction Rate s_cons = α·NLI_entail + β·KP_cover + γ·Rule_cons, where α + β + γ = 1; r_contra: Given comprehensively by the NLI_contradiction rate and the rule conflict ratio; Recommended thresholds: θ_c∈[0.75,0.90], θ_r∈[0.00,0.12], r_cover_min∈[0.70,0.85].

[0027] (2) Agent integration and arbitration The total score S = Σ w_a·score_a, where Σ w_a = 1; w_a can be learned or preset (e.g., compliance has a higher weight). Conflict detection: variance > σ0, any key agent < τ_k, and the set of conflicting opinions is not empty; Arbitration strategy: veto power or Shapley / Bayes weighted average are optional; if the decision is not passed, a rewrite fallback will be implemented.

[0028] (3) Causality and Counterfactual The variable set can be expanded according to the characteristics of the course; The objective function Obj takes into account ΔS (learning gain), Fair (fairness cost), Work (teacher workload), and Time (teaching duration). Counterfactual assessments can employ structural equation modeling, potential outcome frameworks, or simulations. Output intervention u*, expected ΔS, Fair / Work / Time cost, and confidence interval.

[0029] (4) Evidence Chain EC Fields: {doc_id, span, sim∈[0,1], ts, version, source_hash,contributor_agent}; It binds evidence to content, supporting the external display of the minimum verifiable set and the internal auditing of complete evidence.

[0030] (5) Context Protocol CTX Core domains: course metadata, knowledge point mapping, grading details / rubric ID, profile summary, EC index, version and watermark, privacy tags; Operations: append-only logs and audit signatures, supporting cross-session migration and revocation.

[0031] (III) Beneficial Effects Compared with the prior art, the present invention has the following substantial progress and beneficial effects: (1) Strong verifiability and traceability By using verifiable generation (VGE) with evidence chain binding and thresholding for consistency, contradiction rate, and coverage, we ensure that generated materials such as questions, answers, scoring rules, handouts, and feedback are "evidence-based, verifiable, and reversible." When thresholds are not met, automatic rewriting and secondary retrieval are triggered, significantly reducing the problem of inconsistencies between illusion and fact.

[0032] (2) Trustworthy collaborative decision-making among multi-expert agents By introducing a multi-agent collaboration and conflict arbitration mechanism involving question setting, difficulty assessment, grading, compliance, teaching methodology, and arbitration, a unified conclusion that is explainable and applicable can be reached. The veto power of a compliance agent reduces the risk of unauthorized output; If there are significant discrepancies in the scores, the system will automatically enter a closed loop of "diagnosis-rewriting-re-evaluation" to improve the final quality and consistency.

[0033] (3) Causal optimization for learning gain and fairness By combining causal structure learning with counterfactual evaluation, the best intervention among various feasible options is selected, transforming decision-making from "relevance experience" to "maximizing causal gain." Educational equity, teacher workload, and teaching time are incorporated as explicit constraints / costs into the objective function, achieving a globally optimal or suboptimal robust solution under real constraints.

[0034] (4) Contextual consistency and project feasibility With contextual protocols (MCP-type) as the core, course semantics, knowledge point mapping, rubrics and evidence chains are shared across stages. Under the compliance framework of edge-cloud dual ring, low-latency feature processing and cloud-based heavy inference / auditing are achieved. Full-link version watermarking and audit traceability improve the credibility of the project and the visibility of the regulatory authorities.

[0035] (5) Significant systematicity and scalability This invention is not a functional stacking, but rather a combination of technical mechanisms centered on VGE—consistency control—multi-agent arbitration—causal optimization—contextual protocol—dual-loop compliance. This combination can be ported to different educational levels, subjects, and exam types, and seamlessly integrates with existing LMS / academic affairs / question bank systems, possessing excellent scalability and industrialization value. Attached Figure Description

[0036] Figure 1 System overall architecture diagram (data / profile—knowledge—VGE—consistency—multi-agent—causality—protocol—dual loop—compliance).

[0037] Figure 2 Verifiable flowchart for evidence generation and chain of evidence verification (retrieval → generation → chain of evidence → consistency / contradiction rate → rollback / pass).

[0038] Figure 3Multi-expert intelligent agent collaboration and arbitration mechanism (scoring, conflict detection, voting / weighting, rollback and rewriting and finalization).

[0039] Figure 4 Causal structure learning and counterfactual optimization process (causal graph learning → candidate intervention → counterfactual evaluation → objective function optimization).

[0040] Figure 5 Context Protocol (MCP class) and version management relationship diagram (context schema, version watermark, audit backtracking).

[0041] Figure 6 Edge-Cloud Dual-Loop Control Timing Diagram (Classroom Side Low Latency → Cloud Side Inference / Audit → Write-back Update).

[0042] Figure 7 Knowledge graph / scoring rubric / question consistency verification relationship diagram (knowledge point mapping and coverage measurement).

[0043] Figure 8 Example A: Diagram of the closed loop of question setting, grading, reporting, and intervention.

[0044] Figure 9 Example B: Schematic diagram of lecture note generation - consistency verification - multi-agent review - lesson plan / calendar linkage.

[0045] Figure 10 Example C: Classroom Status Access—Profile Update—Counterfactual A / B—Write Back and Scroll Optimization Diagram. Detailed Implementation

[0046] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the specific embodiments of this invention are further described below in conjunction with the accompanying drawings and examples. The following embodiments are for illustrative purposes only and do not constitute a limitation on the scope of protection of this invention. Equivalent substitutions or modifications made by those skilled in the art without departing from the spirit and substance of this invention should fall within the scope of protection of this invention.

[0047] 1. Definition of nouns and variables For ease of understanding, some terms and symbols are defined by convention (see [link]). Figure 1 ): Evidence Chain (EC): refers to the set of supporting evidence for each generated product, containing fields {doc_id, span, sim, source_url / hash, time, version, contributor_agent}, used to achieve a one-to-one correspondence and traceability between generated content and source support.

[0048] Consistency score s_cons ∈[0,1]: The overall consistency index obtained by combining textual implication / contradiction detection, knowledge point alignment and support coverage.

[0049] Conflict rate r_contra ∈[0,1]: The proportion of contradictions between the generated text and the evidence fragments.

[0050] Coverage r_cover ∈[0,1]: The proportion of evidence that covers the key assertions in the generated content.

[0051] Thresholds θ_c, θ_r, and r_min: thresholds for determining consistency, conflict rate, and coverage, with typical ranges of θ_c ∈ [0.70, 0.90], θ_r ∈ [0.00, 0.15], and r_min ∈ [0.60, 0.85].

[0052] A set of multiple agents A: A = {ItemGen (question generator), Difficulty (difficulty rating), Grader (grading), Compliance (compliance), Pedagogy (teaching methodology), Adjudicator (arbitrator)}.

[0053] Weighted voting weight w_a: Σw_a=1, which can be learned or preset and is used for the fusion of scores from various agents.

[0054] Cause-effect graph G=(V,E): V is the set of variables (such as attention ATTN, homework completion rate HW_RATE, question type fit, score SCORE, fairness, etc.), and E is the set of directed edges.

[0055] Objective function: Obj=ΔS−λ1·Fair−λ2·Work−λ3·Time, where ΔS is the counterfactual learning gain, Fair is the fairness cost, Work is the teacher's workload, Time is the time cost, and λ1, λ2, λ3∈[0,1].

[0056] 2 System Overall Architecture and Module Composition like Figure 1 As shown, the overall system architecture of the present invention includes: (1) Data access and learning profile module: Collect multi-source data such as classroom, homework, assessment, course outline and knowledge base to form learning profile (anonymous or de-identified).

[0057] (2) Knowledge retrieval and knowledge graph module: Establish a knowledge graph of "course objectives - knowledge points - questions - scoring rules - reference materials", and realize Top-k evidence recall and knowledge point alignment based on the fusion of vector retrieval and keyword retrieval.

[0058] (3) Verifiable Generation Module (VGE): Generates questions, answers, scoring rules and handouts / feedback in a controlled context, and synchronously outputs a chain of evidence that is bound to each generated content.

[0059] (4) Consistency / Contradiction Rate / Coverage Validation Module: Calculates s_cons, r_contra, and r_cover, and triggers rollback rewriting and secondary retrieval for candidate content that does not meet the threshold.

[0060] (5) Multi-expert agent collaboration and arbitration module: Each expert agent independently scores and interprets, and performs conflict detection, weighted / voting fusion and arbitration.

[0061] (6) Causal structure learning and counterfactual optimization module: learn the causal graph of the relationship between teaching intervention and learning effectiveness, conduct counterfactual evaluation of candidate interventions and select the optimal intervention u* according to the objective function.

[0062] (7) Context Protocol (MCP class) and Version Management Module: Version persistence of course metadata, knowledge point mapping, scoring rules identifier, profile summary, EC index, etc., and generation of version watermark and audit entries.

[0063] (8) Edge-Cloud Dual-Loop Collaboration and Compliance Module: Low-latency feature extraction and PII de-identification are performed on the edge side; retrieval, generation, verification, causal inference and auditing are performed on the cloud side; forming a closed loop of "downlink strategy - edge execution - effect writeback".

[0064] 3. Data Structures and Context Protocols (MCP Class) (1) Context schema The context of this invention adopts a stable and versionable structure, and the core fields include: 1) context_id: A unique identifier for the context, used to associate generated objects with audit records.

[0065] 2) course_meta: Course metadata, which includes at least the course number, chapter / unit identifier, and target learning outcome (LO).

[0066] 3) knowledge_map: A list of knowledge point mappings that records the knowledge point identifier and information such as mastery level / ability level.

[0067] 4) assessment_policy: Reference information for assessment and scoring details, including scoring details identifiers and main evaluation dimensions.

[0068] 5) learner_profile: Learning profile summary, recording the main features of the profile after de-identification (such as attention, task completion rate and other statistics).

[0069] 6) evidence_chain (EC): Evidence chain index, which includes at least the evidence source identifier, the location of the evidence fragment, the similarity or relevance index, and the version number.

[0070] 7) Versioning: Version management information, records timestamps and version watermarks, used for non-repudiation and backtracking.

[0071] 8) Privacy: Privacy and compliance flags, at least record whether it has been de-identified and the authorization / consent status.

[0072] Persistence and Auditing: The context is persisted using a "key fields immutable + version watermark" method; any addition or update to the context by any module automatically generates an audit entry, supporting playback and comparison by version chain (see...). Figure 5 ).

[0073] (2) Chain of Evidence (EC) Fields and Constraints To ensure the verifiability and traceability of the generated content, the Evidence Chain (EC) uses a standardized set of fields for recording and verification (see [link]). Figure 2 , Figure 5 , Figure 7 The core entries are as follows: (a) Required fields 1) doc_id: A unique identifier for the source of evidence (database primary key or external document number).

[0074] 2) span: Evidence fragment location information (page number / line number / paragraph or character range), used to accurately correspond to key assertions.

[0075] 3) sim: Similarity or relevance index, with a value range of [0,1], used to measure the strength of evidence and assertions.

[0076] 4) ts (time): Evidence collection or citation timestamp, used for time-series auditing and "new and old versions" determination.

[0077] 5) ver (version): The version number of the evidence source or index item, used for version playback and difference comparison.

[0078] 6) source_hash: A hash digest of the source content or metadata, used for integrity verification and non-repudiation.

[0079] 7) contributor_agent: The module / Agent identifier that generates or selects this evidence (e.g., “Retriever”, “VGE”, “Grader”).

[0080] 8) binding_index: A one-to-one index corresponding to the key assertions in the generated content (e.g., "assertion ID" or "paragraph-sentence number").

[0081] (ii) Optional fields 9) Confidence: Confidence value of evidence reliability ([0,1]), which is different from sim and reflects the credibility of the source or the level of review.

[0082] 10) policy_tag: compliance / license tag (copyright, open license, scope of use), used for export control.

[0083] 11) Provenance: Source path or crawling link information (such as collector ID, crawling time, preprocessing pipeline identifier).

[0084] 12) anchor_kp: The set of nodes bound to the knowledge graph (KP identifier, level, capability level) to facilitate coverage calculation.

[0085] 13) locale / units: Language and unit of measurement information, used for cross-language and cross-unit consistency verification.

[0086] (III) Structural and Consistency Constraints 14) One-to-one binding: Each "key assertion" must be bound to at least one EC record; the same evidence can support multiple assertions, but binding_index must be recorded separately.

[0087] 15) Integrity verification: The referenced doc_id and span must be able to be replayed and located; source_hash is used to compare the source content to ensure it has not been tampered with.

[0088] 16) Measuring reasonableness: When sim or confidence falls below a threshold (e.g., 0.6 or set according to the scenario), a rollback re-examination or supplementary evidence is triggered; it is linked with consistency metrics s_cons, conflict rate r_contra, and coverage rate r_cover to determine whether it passes (see...). Figure 2 , Figure 7 ).

[0089] 17) Version Consistency: EC entries are updated synchronously with the context versioning; changes require the generation of a new version and the retention of old version snapshots, supporting replay by version chain (see [link]). Figure 5 ).

[0090] (iv) Lifecycle and Audit 18) Write strategy: EC writes immediately during the generation phase; any modification or replacement adds a new audit entry and does not overwrite the history.

[0091] 19) Audit Field Linkage: Audit records must include at least ctx_id, agent_trace, metrics (s_cons, r_contra, r_cover), decision (pass / rollback), version_hash, and signature; EC and audit entries are linked via context_id and binding_index (see...). Figure 5 ).

[0092] 20) Minimum Verifiable Set: Outputs a necessary subset of EC (masking sensitive information) when displayed externally, while retaining all entries internally to support accountability tracking and comparison playback.

[0093] (v) Privacy and Compliance 21) De-identification: ECs must not contain personally identifiable information; if the source of evidence involves personal data, it should be recorded in an anonymized extract or summary and labeled with a policy_tag.

[0094] 22) Withdrawal and Clearance: When the right to be forgotten is triggered, the identifiable link is deleted while the irreversible statistical characteristics are retained; if the deletion causes the binding_index to break, the consistency index should be reassessed in the "evidence missing" state and a prompt should be made to supplement the evidence.

[0095] (vi) Compatibility and Expansion 23) Cross-model compatibility: EC field naming and value range are kept compatible, allowing the addition of new fields without breaking existing parsing; supports locale tags from multiple language sources and units tags for unit conversion.

[0096] 24) Index acceleration: Create a composite index for doc_id, binding_index and anchor_kp to support bidirectional retrieval and batch verification of "assertion → evidence" and "evidence → assertion".

[0097] 4 Verifiable Generative Evidence (VGE) and Chain of Evidence Verification Process (1) Generation and rollback process See Figure 2 The process includes: S301 Task Analysis: Construct contextual requirements based on the task and search query, and determine the knowledge points, difficulty, and format constraints; S302 Evidence Retrieval and Alignment: Perform fusion retrieval and KP mapping in the knowledge base / graph to obtain Top-k evidence; S303 Controlled Generation: Candidate content (questions / answers / scoring rules / handouts, etc.) is generated by the generative model under the constraints of the aforementioned context. S304 Evidence Chain Binding: Establishing a one-to-one correspondence between candidate content key assertions and retrieved evidence (EC). S305 Consistency metrics: Calculate s_cons, r_contra, and r_cover; S306 Threshold Judgment and Rollback: When s_cons < θ_c or r_contra > θ_r or r_cover < r_min, perform controlled rewriting and secondary retrieval, and return to S303; S307 Output and Audit: Output the finalized content and its chain of evidence, and write it to the context protocol and audit log.

[0098] (2) Key points of indicator calculation and consistency verification Textual implication / contradiction detection: Using an NLI model or a rule-statistical hybrid method, s_cons and r_contra are obtained.

[0099] Knowledge point coverage: Based on the mapping between the question stem / answer / scoring details and the knowledge graph, r_cover is calculated and a “KP×Item” coverage matrix is ​​output for gap diagnosis.

[0100] Numerical / structural consistency: Verify the formulas, units, boundary conditions, and the cross-references between the question stem, answer, and scoring criteria.

[0101] Typical thresholds: θ_c∈[0.70,0.90], θ_r∈[0.00,0.15], r_min∈[0.60,0.85].

[0102] 5. Multi-expert intelligent agent collaboration and arbitration (1) Scoring fusion and conflict detection See Figure 3 The steps are as follows: S401 Rating Collection: Agents such as ItemGen, Difficulty, Grader, Compliance, and Pedagogy each rate the candidate content in the range [0,1] and provide reasons; S402 Weighted Fusion: Calculate the total score S based on preset or learnable weights w_a (Σw_a=1); S403 Conflict Detection: A conflict is identified when the score variance > σ0 or the key agent score < τ_k; S404 Arbitration and Rollback: If the compliance agent rejects the application or there is a conflict, a targeted rewrite will be triggered (such as reducing the difficulty, supplementing evidence, rearranging the structure), and the application will be returned to S401 for review. S405 Finalization and Audit: Upon approval, the finalized version content, scoring instructions, and audit log will be output.

[0103] (2) Arbitration Rules and Parameters Compliance agents have veto power; Low scores for both question design and difficulty triggered a "rewrite with reduced difficulty"; When the teaching methodology score is high but conflicts with the test score, priority should be given to ensuring consistency with the teaching objectives; Typical parameter range: τ∈[0.70,0.90], σ0 is set empirically based on historical score variance.

[0104] 6. Causal Structure Learning and Counterfactual Optimization like Figure 4 As shown, this module includes: S501 Causal Graph Learning: Learning causal graph G=(V,E) based on profile, historical interaction and prior teaching structure; S502 Candidate Intervention Generation: Forming an intervention set U including question quantity / question type / order / lecture handout structure / pacing / differentiated assignments, etc. S503 Counterfactual evaluation: For each candidate u∈U, perform counterfactual inference on G to obtain y_cf; S504 Objective function optimization: Calculate Obj=ΔS−λ1·Fair−λ2·Work−λ3·Time, and select the optimal intervention u* under the constraints of workload and duration; S505 Write and Distribute: Write u* and expected gain to the context (including version watermark and audit) and synchronize to the lesson plan / teaching calendar / learning task sheet.

[0105] 7 Context Protocol and Version Management See Figure 5 : Version semantics: Use major.minor.patch to manage compatible and incompatible changes; Version watermark: version_hash=H(core field + timestamp + signature), and is signed with a private key and verified with a public key to achieve non-repudiation; Version control: create, update, archive and rollback, supports read-only snapshots; Branching and Merging: Merging strategies for multi-task / multi-model parallel contexts to resolve field conflicts and priorities; Minimum verifiable set: Only necessary evidence and summaries are made available to the public, while the full set is retained internally for audit backtracking.

[0106] 8 Edge-Cloud Dual-Loop Control and Compliance like Figure 6 As shown: Edge side (classroom end): Performs data collection and lightweight feature extraction, PII de-identification, and only uploads anonymized features or statistics; optionally, it provides local fast feedback.

[0107] Cloud side: Performs data access / decryption, knowledge retrieval, VGE, consistency verification, multi-agent arbitration, causal optimization, context writing, and audit archiving.

[0108] Downlink and Writeback: The cloud side outputs u* and task orders to the edge for execution; after classroom implementation, the effect is collected and the profile is written back, forming a double-loop closed loop of "downlink - execution - writeback - re-optimization".

[0109] Auditing and Recall: Audit entries must include at least ctx_id, agent_trace, EC, indicator value, decision result, version_hash, and signature; when the right to be forgotten is triggered, context-level clearing is supported while retaining necessary irreversible statistical characteristics.

[0110] Example 1: Question creation—grading—reporting—intervention closed loop like Figure 8 As shown: (1) Input course objectives, knowledge points and target difficulty (mean μ_d, variance σ_d); (2) The system retrieves textbooks and question banks, VGE generates test questions and scoring rules, and the evidence chain coverage r_cover ≥ preset lower limit; (3) If the consistency and contradiction rate meet the threshold, the system will automatically roll back and rewrite. (4) Multi-Agent scoring and arbitration: If the interpretability of the step is below the threshold, the "refine step comments" rewrite is triggered; (5) After the exam is conducted, the system will automatically grade the papers and generate multi-dimensional reports (average score, distribution, accuracy rate of knowledge points, and common mistakes). (6) The causal module performs counterfactual evaluation on candidate interventions (such as "increasing example explanations", "lowering the upper limit of the difficulty of the next test", "differentiated homework A / B sets"), and selects u* by Obj; (7) Synchronize u* and the summary to the lesson plan and teaching calendar and archive and audit them.

[0111] Example 2: Lecture Notes Generation—Consistency Verification—Multi-Agent Review—Lesson Plan / Calendar Linkage like Figure 9 As shown: (1) Input chapter objectives and weakness profiles; (2) VGE generates a draft of lecture notes with the structure of "example problem - breakdown - comparison - misconception - summary - consolidation problem" and binds it to EC; (3) Calculate s_cons, r_contra, and r_cover. If the threshold is not reached, roll back and rewrite. (4) Multi-Agent review: Pedagogy score high, Compliance pass, Difficulty suggests lowering the difficulty of the comprehensive questions; (5) Write the handouts and assignment list into the context and publish them to the lesson plan / teaching calendar (automatically reserve time for Q&A); (6) Version management and watermarking for subsequent audit playback.

[0112] Example 3: Classroom Status Access—Profile Update—Counterfactual A / B—Write-back and Scrolling Optimization like Figure 10 As shown: (1) At the edge, attention, interaction frequency, etc. are anonymized and the minimum necessary uplink is performed; (2) The portrait update triggers the re-evaluation of the causal graph, and identifies the causal chain of "explanation speed - reading comprehension - error cause"; (3) Propose A / B intervention: A "Explanation speed -10%", B "Test frequency +1 / week"; (4) Counterfactual assessment shows that Option A has a higher ΔS and Option B has a lower workload; (5) After multiple agents and teachers make decisions on u*, the observation data is written back one week later and the cause-effect graph and strategy library are continuously optimized.

[0113] This invention can be deployed in K-12, higher education, and vocational education settings, and is suitable for online / offline / blended learning. Through standardized context protocols and APIs, it can interface with academic affairs systems, learning management systems (LMS), electronic question banks, and digital textbook platforms.

[0114] This specification provides a complete system architecture, process methods, key algorithms, data structure definitions, threshold ranges, and implementation examples. For those skilled in the art, any equivalent substitutions or modifications made to module implementation details, parameter values, number and names of agents, causal learning algorithms, context protocol formats, encryption and privacy strategies, etc., without departing from the spirit and substance of this invention, should fall within the protection scope of this invention.

Claims

1. A verifiable generative and causal optimization-oriented intelligent educational system for integrated teaching and assessment, characterized in that, The system includes: a data access and learning profile module, used to collect classroom data, homework data, assessment data, and course syllabus data, generating learning profiles that include behavioral characteristics and knowledge point mastery; a knowledge retrieval and knowledge graph module, used to establish a relationship graph between course objectives, knowledge points, questions, scoring rules, and reference materials, and to achieve evidence retrieval and knowledge point alignment based on vector retrieval and keyword retrieval, sorted by relevance; a verifiable generation module, used to generate questions, answers, scoring rules, and teaching materials or feedback under the constraints of retrieval results, and synchronously output a chain of evidence corresponding to the generated content; and a consistency and contradiction rate verification module, used to calculate the consistency score, contradiction rate, and evidence coverage rate between the generated content and the evidence chain, and to trigger a rollback and re-verification when any indicator fails to meet the threshold condition. The system includes: a multi-expert collaboration and arbitration module, comprising a question generation module, a difficulty labeling module, a grading module, a compliance module, a pedagogy module, and an arbitration module, used for scoring and fusing candidate outputs, conflict detection, and arbitration decisions; a causal structure learning and counterfactual optimization module, used to learn the causal relationship between teaching interventions and learning outcomes, perform counterfactual evaluations of candidate interventions, and select the optimal intervention according to the objective function; a context protocol module, used to version-store course metadata, knowledge point mappings, scoring rubric identifiers, learning profile summaries, and evidence chain indexes, and generate version watermarks and audit logs; and an edge and cloud collaboration and compliance module, used to perform low-latency feature extraction and de-identification processing on the edge side, and to perform retrieval, generation, verification, causal inference, and audit archiving on the cloud side, and supports result write-back and closed-loop optimization.

2. An educational intelligent method based on the system described in claim 1, characterized in that, include: S1 collects multi-source data and builds learning profiles; S2 retrieves task-related evidence from the knowledge base and graph, aligns knowledge points, and forms context; S3, the verifiable generation module generates candidate content under context constraints and simultaneously constructs an evidence chain; S4 calculates consistency score, contradiction rate, and coverage rate. If the threshold conditions are not met, a rollback and rewriting are performed until the threshold is met or the process terminates; S5, the multi-expert collaboration and arbitration module performs score fusion and conflict arbitration. If the conflict fails, a targeted rewriting and re-evaluation are triggered; S6 learns the causal structure based on the learning profile and historical data, performs counterfactual evaluation on candidate interventions, and selects the optimal intervention based on the objective function; S7 writes version watermarks and audit records through the context protocol and synchronizes the content and intervention plan to lesson plans, teaching calendars, and learning task sheets; S8 deploys and executes under an edge and cloud collaborative architecture, and collects implementation effect feedback to continuously update the learning profile and causal structure.

3. A computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the method of claim 2.

4. The system or method according to claim 1 or 2, characterized in that, The threshold conditions set by the consistency and contradiction rate verification module include: the consistency score is greater than or equal to the first threshold, the first threshold ranges from 0.70 to 0.90; the contradiction rate is less than or equal to the second threshold, the second threshold ranges from 0.00 to 0.15; the evidence coverage rate is greater than or equal to the third threshold, the third threshold ranges from 0.60 to 0.85; when any condition is not met, a second retrieval, controlled rewriting, and recalculation of the indicators are performed, and the number of rollbacks and the reasons are recorded in the audit entry.

5. The system or method according to claim 1 or 2, characterized in that, The evidence chain includes at least several of the following fields: document identifier, fragment location information, similarity index, timestamp, version number, source hash summary, and contribution module identifier; the evidence chain and the generated content are bound one-to-one through an index to provide a minimum verifiable set externally and to conduct full audit internally.

6. The system or method according to claim 1 or 2, characterized in that, The multi-expert collaboration and arbitration module employs a weighted fusion and conflict detection mechanism, satisfying the following rules: the compliance module has veto power; when the scores of the question-setting module and the difficulty calibration module are both below their respective thresholds, a rewrite with reduced difficulty is triggered; when there is a conflict between the scores of the teaching methodology module and the question-setting module, the consistency of teaching objectives is prioritized; the total score is determined by the weighted sum of the scores of each module and their corresponding weights, with the sum of the weights being one; when the score variance exceeds the preset upper limit or the score of a key module is below the threshold, the arbitration rollback process is initiated.

7. The system or method according to claim 1 or 2, characterized in that, The causal structure learning and counterfactual optimization module learns causal structures based on a set of variables including attention, homework completion rate, reading ability, test frequency, difficulty level, score and fairness indicators, and selects the best one using an objective function. The objective function is the learning gain minus the weighted sum of fairness cost, teacher workload cost and time cost. The weight coefficients range from zero to one and satisfy the upper limit of teacher workload and upper limit of teaching time.

8. The system or method according to claim 1 or 2, characterized in that, The context protocol module adopts a versioned structure, including course metadata, knowledge point mapping, scoring rules identifier, learning profile summary, evidence chain index and privacy tag field, and generates version watermark by hashing core fields, timestamps and signature information; version operations include at least creation, update, archiving and rollback, and support cross-model and cross-session context compatibility and playback.

9. The system or method according to claim 1 or 2, characterized in that, The edge and cloud collaboration and compliance module performs personal identification information de-identification and minimum necessary data uplink on the edge side, and performs retrieval, generation, verification, causal inference and audit archiving on the cloud side. Audit entries record at least the context identifier, module processing trajectory, evidence chain, values ​​of each indicator, decision results, version watermark and signature, and support version-based backtracking and responsibility tracing.

Images