Live broadcast interaction-oriented arrangement and compliance law enforcement method and system, and storage medium

Abstract

The invention discloses a live broadcast interaction-oriented arrangement and compliance law enforcement method and system, and a storage medium. According to the method, an intelligent LLM cluster is used as a core execution unit, live broadcast interaction intermediate representation (LiveFlow-IR) represented by a directed acyclic graph (DAG) is adopted, and integrated control of orchestrable and law enforcement is carried out on the whole process from collection, governance, semantic restoration, clustering, clarification, distribution, reply / rejection to audit trace leaving. The system realizes real-time complementation, understanding and commissioning / conversion in a multi-language, multi-dialect, discontinuous, broken-support, unclear, incoherent and incomplete-semantic input scene; comprehensive decision making is carried out on the answerability, the commercial value, the compliance risk and the time delay cost through multi-target scoring, and a direct answer / clarification / rejection path is selected; before output, calculating a compliance criterion vector and executing compliance gating, and when a threshold value is not reached, degrading into a conservative verbal skill or refusing to answer and giving a substitution suggestion; outputting a verifiable evidence formed by binding with the data, and carrying out committed trace leaving; sLO (Service Level Object) steady-state operation is realized by combining risk budget, back pressure and drift control; and orderly output of a single foreground image is ensured through a speaking right Token ring. Compared with an existing loose integration scheme, the method gives consideration to interaction quality, compliance verifiability and time delay stability in a high-concurrency and strong-uncertainty environment, and is suitable for various live broadcast scenes such as e-commerce, education, government affairs and text travel.

Description

Technical Field

[0001] This invention relates to the field of intelligent interaction and compliance auditing technology, specifically to a live interactive orchestration and compliance enforcement method, system and storage medium based on intermediate representation (LiveFlow-IR) and intelligent agent LLM cluster, belonging to the cross-technical direction of real-time human-computer interaction, content security and verifiable auditing. Background Technology

[0002] Today's live streams are no longer just "one-way broadcasts." Viewers ask questions as they watch, and the chat is filled with colloquialisms, typos, dialects, and incomplete sentences, as well as spamming, attention-seeking, and sensitive topics. Traditional methods either rely on human hosts to handle things on the spot or use "talking digital humans" to read lines from a script. The former is difficult to guarantee in terms of efficiency and compliance, while the latter lacks genuine interaction. Simply piling up ASR, LLM, and TTS is not enough: the problem lies in the lack of process orchestration and enforcement—when should clarification be given before answering? How to automatically close the conversation when encountering high-risk questions? How to stabilize P95 latency during peak hours? How can the answers contain "evidence" for easy explanation afterward? All of these require an executable and auditable "path" to coordinate. We propose abstracting the entire live stream into an "executable graph" (LiveFlow-IR), with clear goals and constraints for each step, and a scheduler to execute it; before each output, "causal guardrails" provide evidence for the answer, and the summary is made into an irreversible commitment. In this way, live streaming becomes more than just a talking machine; it becomes a structured execution system that balances interaction, efficiency, and compliance in real time, ensuring smooth operation during peak hours, avoiding sensitive situations, and allowing for the review of responses. This is the key upgrade for the live streaming industry, moving from simply "being able to talk" to "speaking responsibly." Summary of the Invention

[0003] I. Purpose of the Invention This invention aims to construct a programmable, enforceable, and verifiable live streaming interaction path: (1) Propose LiveFlow-IR, which uses DAG to organize nodes such as Ingest / Repair / Cluster / Clarify / Assign / Answer / Refuse / Audit into an executable graph, with SLO and Policy-ID carried on the edges, so that the scheduler executes on IR instead of on scattered modules; (2) Optimize unified interaction quality and operation objectives with multiple objectives: Define answerability A, business value B, compliance risk R, and delay cost L to form a score function and support path-level shifting (clarify first and then answer directly or refuse to answer directly). (3) Introduce causal guardrails before output and calculate compliance justification vector E (topic consistency, sensitive domain, source credibility, evidence existence). If the threshold is not reached, it will be downgraded to conservative wording / refusal to answer + alternative suggestions. (4) The answer is bound to E through Proof-Carrying Output and written into the audit with hash / tree commitment, which supports post-event verification; (5) Design risk budget R_total and back pressure / drift control. When the budget is insufficient, there is peak congestion or model abnormality, SLO degradation and priority reordering are triggered to ensure overall stability and compliance. (6) The Token ring speaking rights protocol is adopted to ensure that no single front-end user can simultaneously interrupt the conversation; The goal is to enable live streaming to achieve lower P95 latency, higher one-time resolution rate, lower violation rate, and stronger auditability in high-concurrency, highly uncertain, and strictly compliant environments, and to achieve adaptive optimization through metric write-back in multiple sessions.

[0004] II. Technical Solution

[0005] (I) Introduction This invention discloses a technical framework for orchestration and compliance enforcement in live streaming interaction. The system uses LiveFlow-IR as an intermediate representation, organizing Ingest, Repair, Cluster, Clarify, Assign, Answer, Refuse, and Audit nodes into a Directed Acyclic Graph (DAG), with SLOs and Policy-IDs bound to the edges. During runtime, the system is driven by an IR compilation and execution engine, performing path selection and shifting based on score = α·A + β·B − γ·R − δ·L. Before output, a causal guardrail calculates the E vector and executes a Compliance Gate. The output and E together constitute the Proof-Carrying Output, which is written to the audit storage via hash / tree commitment. The system introduces a risk budget R_total, backpressure / drift detection, and SLO degradation strategies, while using a Token ring protocol to ensure orderly speech by a single front-end user. This framework enables full-process visualization, verification, and controllability of live streaming interaction, from acquisition, repair, aggregation to response / rejection and auditing.

[0006] (II) Methods and Procedures

[0007] S0 initialization Read the session configuration and operating parameters, including at least the Service Level Objective (SLO), compliance policy library, risk budget cap (R_total), speaking rights configuration, weight parameters α / β / γ / δ, window length W, and threshold set {θ_score, θ_E, θ_u, θ_i}. Establish the registry for the agent LLM cluster, load role metadata and policy mapping relationships, and initialize the audit channel and commitment calculator.

[0008] S1 LiveFlow-IR Construction Construct an intermediate representation, LiveFlow-IR, using a directed acyclic graph (DAG). The node set includes at least Ingest, Repair, Cluster, Clarify, Assign, Answer, Refuse, and Audit. Bind SLOs and Policy-IDs to edges and record optional priorities, timeouts, and failure fallback paths. IR allows for local shifting and rerouting based on load and risk during runtime.

[0009] S2 Ingestion The system receives bullet comments and live chat input, forming a time window stream. It performs edge-side speech activity detection and multilingual / dialect recognition on the voice input, splicing and transing discontinuous speech segments into text segments and a confidence stream. After aligning the timestamps of the text and speech, they are grouped into the same event stream, generating an input object containing source, language, dialect, and preliminary quality labels.

[0010] S3 Treatment Pretreatment The input objects undergo deduplication, screen-folding, and civilized rewriting and downgrading of profanity and inflammatory language, generating governance tags (such as violation risk, repetition rate, and sentiment level). The governance results are appended to the input objects for subsequent sorting and enforcement.

[0011] S4 Semantic Repair Based on an agent-based LLM cluster, entity completion, word order rewriting, ambiguity resolution, and dialect word mapping are performed on discontinuous, ambiguous, incomplete, and dialect-mixed text. The output is the repaired text, uncertainty u, and completeness i. Optionally, a "paraphrased / retelled" version can be created for subsequent clarification and restatement. (Optional formulas, including but not limited to) u = 1 − comp, i = cov · coh Where comp represents content completeness estimation, cov represents key entity coverage, and coh represents syntactic coherence measurement.

[0012] S5 Clustering Within a sliding window W, the repaired text is aggregated based on semantic vectors and key entities to generate several "standard questions" and their representative expressions. Support and diversity indices are calculated for each cluster to form a cluster summary.

[0013] S6 Multi-Objective Scoring and Routing Calculate a multi-objective score for each standard question and determine the path. Scoring factors should include at least responsiveness (A), business value (B), compliance risk (R), and time delay cost (L). (Scoring formulas, including but not limited to) score = α·A + β·B − γ·R − δ·L When score ≥ θ_score and u and i meet the thresholds, the system enters the direct response path; when u < θ_u or i < θ_i, the system enters the clarification path; when R is high or the Policy-ID indicates a restricted topic, the system enters the rejection path. Scoring and routing results are recorded in the IR execution track.

[0014] S7 Clarify (optional) For questions with low completeness or high ambiguity, generate clarification cards containing at least two candidate question types or key entity completion items. Update weights or thresholds based on user selections and feedback, forming a local adaptive mechanism.

[0015] S8 Assignment and Speaking Rights Control The routing results are dispatched to the response generation unit or the rejection generation unit, and the sub-agent in the agent LLM cluster that matches the current Policy-ID is invoked to generate the content. A speaking rights token ring protocol is used to pass speaking rights among multiple agents, ensuring that only one image is output at the front end. Time slices and timeout fallback policies are set for tasks that are currently outputting.

[0016] S9 Cause and Effect Guardrail Enforcement Before generating audible / visual output, calculate the compliance justification vector E, which includes at least the topic consistency score, sensitivity area score, source credibility score, and evidence existence indicator. Implement compliance gating based on the threshold θ_E; if the threshold is not met, downgrade to conservative wording or refuse to answer, and provide alternative suggestions and authoritative channels. (Gating criteria, including but not limited to) gate = 1 if (w·E ≥ θ_E); otherwise gate = 0 Where w is the policy-related weight vector.

[0017] S10 Proof-Carrying Output The content to be output is bound to the E vector to form a rationalized output record. A commitment value is calculated for auditing purposes. (Commitment calculation, including but not limited to) H = Hash(digest|| E || timestamp|| random number).

[0018] S11 Audit Recording The audit store writes a summary and commitment value of "Issue → Policy → Output → Rationale," supporting retrieval by session, time, or Policy-ID index. Audit records are stored with minimal requirements, excluding sensitive fields that could identify personal information.

[0019] S12 Risk Budgeting and Adaptation Maintain the session risk budget R_total and remaining budget R_remain. Deduct the budget based on the risk score for each response and recharge it upon positive compliance feedback. Monitor back pressure and drift, and implement capping and SLO downgrading as necessary. (Budget updates and triggering conditions, including but not limited to) R_remain = R_remain − η·R + κ·refund Backpressure trigger conditions: L_P95 > SLO, or queue length Q > Q_max, or throughput T < T_min Triggering drift condition: consistency difference Δc ≥ τ.

[0020] S13 indicator write-back Metrics such as P95 latency, first-time resolution rate, violation trigger rate, and audit hit rate are collected, written back to the parameter library, and α / β / γ / δ and thresholds are updated online. The update strategy does not involve training details and mainly focuses on parameter adjustments driven by indicator signals.

[0021] (III) System Module Description (1) IR builder Used for generating and editing LiveFlow-IR, supporting the addition of nodes / edges, binding of SLO and Policy-ID, setting of fallback paths and priorities, and providing versioning and hot update capabilities; (2) IR compilation and execution engine Compiles IR into an executable pipeline, supporting path-level shifting, hot-plugging, failure retries, and idempotent control. Provides an event-driven execution model and fine-grained timing to meet SLO requirements; (3) Input acquisition module (Ingest) Unified access to bullet comments and live chat input, completes voice activity detection, multilingual and dialect recognition, discontinuous voice splicing, timestamp alignment and source labeling, and transcribes voice into text and outputs quality labels; (4) Governance Module Perform deduplication, screen-folding, profanity and extortion detection, civility rewriting and risk labeling, and output governance labels to influence subsequent sorting and law enforcement; (5) Semantic Repair Module Based on an intelligent agent LLM cluster, entity completion, word order rewriting, dialect and spoken language mapping, and paraphrasing / retelling generation are completed; the output is the repaired text, uncertainty u, completeness i, and key entity table. (6) Clustering module Within window W, aggregate semantic vectors and key entities to generate standard questions and representative statements, and calculate support and diversity indices; (7) Scoring and Routing Module Calculate A, B, R, and L to form a score, and perform path selection for direct response / clarification / rejection; combine SLO and governance tags to perform priority rearrangement and fallback; (8) Clarify module Generate a clarification card containing multiple candidate questions or entity completion items, receive feedback from clicks, and update parameters; (9) Scheduling and speaking rights module (Assign + Token Ring Manager) To achieve the distribution of intelligent agents in an LLM cluster, a token ring protocol, time slices, and timeout rollback are used to ensure the orderliness and continuity of the output of a single front-end image. (10) Answer generation module (Answer) / Rejection generation module (Refuse) Generate audible / visual content for both direct and non-response paths, supporting multilingual and multi-dialect expression and narration / retelling, with voice serving only as a means of expression; (11) Causal barrier and evidence generation module (Compliance Gate + Evidence Vector) Calculate the compliance justification vector E and perform gating; automatically downgrade content that does not reach the threshold to conservative wording or refuse to answer, and generate alternative suggestions and channels; (12) Audit and Commit module Generate a commitment value H from the output and E and write it to the audit storage; provide a session- and time-based searchable interface and a verification interface; (13) Risk Budgeting and Adaptive Control Module Maintain R_total and R_remain, monitor back pressure and drift, trigger capping, force priority clarification and SLO degradation; feed back the running signals to the IR execution engine; (14) Strategy library / compliance library / parameter storage Stores Policy-ID, SLO, threshold set, and weight parameters α / β / γ / δ, supporting grayscale and version management; (15) Indicators and Learning Module The system collects metrics such as latency, resolution rate, violation rate, and audit hit rate, and updates them online without involving training details. (16) Multilingual output and facial expression rendering module (optional) Render the text output into multilingual, multi-dialect speech and facial expression / lip movement sequences to ensure the consistency and uniformity of the front-end image; (17) Edge acceleration and caching module (optional) Lightweight inference, caching, and prefetching are performed on the endpoint and edge sides to reduce end-to-end latency.

[0022] III. Security Primitives / Service Invocation Structure and Boundary Protection Mechanism Explanation 1) Security primitives include, but are not limited to: Gate() (compliance gate), Budget() (risk budget read / write), TokenGrant() / TokenRelease() (grant / reclaim speaking rights), Commit() (commit calculation and writing), Route() (path selection), Throttle() (rate limiting and capping), Clarify() (clarification generation), and Refuse() (rejection template). 2) The service call structure adopts a layered and decoupled approach: the front-end rendering layer only receives output with rationale; the orchestration layer executes according to LiveFlow-IR; the compliance and enforcement layer performs Gate() judgment on the E vector; the audit layer generates commitments using Commit(); and each layer exchanges the minimum necessary data through controlled interfaces. 3) Boundary protection mechanisms should include at least: loading blacklists, whitelists, and glossaries by Policy-ID; performing parameter validation and timeout circuit breaking for cross-domain calls; removing sensitive information and watermarking outgoing content; prioritizing rejection and providing alternative suggestions for high-risk topics; and enabling Throttle() and SLO degradation when back pressure and drift are triggered. 4) Permissions and Isolation: The agent LLM cluster serializes speaking rights through TokenGrant(); it is allowed to enter the output phase only when Gate() passes and Budget() is sufficient; all primitive calls are recorded in a minimal audit log and associated with the commitment H to ensure verifiability and traceability.

[0023] IV. Terminology Definitions and Identification Boundaries 1) "LiveFlow-IR" is the intermediate representation expressed by a directed acyclic graph (DAG); "SLO" is the service level target; "P95" is the 95th percentile delay; "Token ring" is the ring-shaped right-to-speak protocol; "E vector" is the compliance justification vector; 2) "Standard Question" refers to the representative question obtained through semantic clustering within a sliding window; "Clarification Card" is an interactive unit containing at least two candidate questions or entity completion items; "Conservative Statement" is the minimum necessary answer and alternative suggestion set output when there is high risk or insufficient evidence. 3) "Risk Budget" is the quantitative upper limit of the compliance risk that can be tolerated for a session; "Back Pressure" is the trigger condition for service degradation due to congestion; "Drift" is an abnormal signal of output confidence or consistency; "Refusal Template" is a structured script containing reasons, terms and conditions, alternative suggestions and channels. 4) The identification boundaries include at least: completion and paraphrasing of multilingual, multi-dialect, and discontinuous or incomplete speech text; topic identification in sensitive fields such as medical and financial fields; governance of spamming, extortion, and adversarial variants; and entering the rejection or clarification path when the E-vector is below the threshold or the budget is insufficient. All of the above identifications are based on operating signals and rules and do not involve training details.

[0024] V. Explanation of Path Control and Abnormal Detour Blocking Mechanism 1) Path control is based on LiveFlow-IR: Route() selects between "direct answer / clarification / rejection" based on score, SLO and governance tags; it supports shifting and rollback at the node level to form a fine-grained execution trajectory; 2) Abnormal detours include at least the following: when Gate() fails or the E vector is insufficient, force a detour to Refuse() or Clarify(); when Q>Q_max or P95 exceeds the threshold, trigger Throttle() and postpone non-critical issues; when drift occurs, pause the relevant agent and switch to a backup model; 3) Blocking mechanisms include, but are not limited to: edge blocking of high-risk Policy-IDs (prohibiting direct answers, only allowing clarification or refusal to answer); window-level folding of low-value questions with high repetition rates; global blocking of "direct answers to sensitive areas" when risk budget is insufficient, and uniform conservative wording; 4) Recovery and rollback: After the load and risk have decreased, the blockade is gradually lifted according to the priority queue and time window to restore the original path; all blockade and detour changes are written to the audit trail via Commit() and bound to the commitment value H for post-event verification.

[0025] VI. Beneficial Effects Compared with the loose integration of "ASR+LLM+TTS", this invention brings the following advantages: (1) Programmable: LiveFlow-IR is used to uniformly describe the process and constraints, the scheduler executes the graph rather than the execution module, and supports adaptive path shifting; (2) Enforceable: The causal guardrail provides an E vector before output, which automatically triggers conservative mode or refuses to answer, reducing compliance incidents; (3) Verifiable: The output and E form a Proof-Carrying Output, and the trace is left with hash / tree commitment, which can be verified afterward without exposing privacy; (4) Stable latency: SLO, risk budget and back pressure work together, P95 is controllable during peak hours, avoiding "microphone jamming" and "fighting sound"; (5) High hit rate: The repair-clustering-clarification link improves the "answerability" and increases the one-time resolution rate; (6) Adaptive: The index is updated online by writing back α / β / γ / δ, and becomes more accurate the more it is used across different sessions.

[0026] Overall, it achieves a dynamic balance between interaction quality, commercial conversion, and compliance risks, providing feasible infrastructure capabilities for various scenarios such as live-streaming e-commerce, education, and government affairs. Attached Figure Description

[0027] Figure 1 System overall block diagram.

[0028] Figure 2 LiveFlow-IR (DAG) example and policy binding graph.

[0029] Figure 3 Sequence diagram of semantic repair and clarification pipeline.

[0030] Figure 4 A diagram illustrating multi-objective scoring and path selection.

[0031] Figure 5 Scheduling and speaking rights token ring sequence diagram.

[0032] Figure 6 Compliance guardrails, justification output, and audit commitment structure diagram.

[0033] Figure 7 Flowchart of risk budget / back pressure / drift and SLO degradation control. Detailed Implementation

[0034] To make the objectives, technical solutions, and beneficial effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. This invention is not limited to the specific embodiments described below; any equivalent substitutions or improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

[0035] Example 1: Basic Scenario In a live-streaming sales event for small home appliances, only a digital human host was speaking on the front end, while a cluster of intelligent agents (LLMs) was collaborating in the back end. It wasn't a voice assistant; its core function was understanding and decision-making, with voice merely serving as a means of expression. It could switch between Chinese, English, and dialects at any time. Viewers commented intermittently, "Is this...induction cooker...okay? Will it burn? How big is it?" Others asked in dialects, "Want one?" The system first completed and rewrote these fragmented / multi-dialect inputs in real time, merging them into standard questions: "Is an induction cooker usable? Is it prone to sticking? What sizes are available?" If it was still unclear, a clarification card would pop up for viewers to select from. The scheduler assigns different roles to intelligent agents based on the pace and risk of the event: the anchor agent gives a conclusion in one sentence; the expert agent adds two key parameters; the after-sales agent explains returns and maintenance clearly in one go; and the atmosphere agent appropriately inserts dialects and bilingual trivia (such as "non-stick = non-stick"). Throughout the process, a speaking token loop ensures that only one voice is output and no one interrupts. As the peak of a major promotion approaches, the system automatically limits traffic and prioritizes clarification to ensure stable P95 latency; when encountering inappropriate or high-risk topics (such as medical efficacy), the compliance barrier is triggered, switching to an explainable refusal to answer and providing official channels. Before each answer is output, a reasoning vector explaining "why this answer is acceptable" is generated. This vector, along with the answer itself, generates a commitment value that is written into the audit, allowing for future verification that the answer has not been tampered with. Each session also has a risk budget: when risk consumption approaches its limit, the system automatically switches to a more conservative approach and strategy. Through this arrangement and enforcement approach, the live stream maintains natural interaction and multilingual / dialectal expression while balancing efficiency, compliance, and traceability in complex, high-concurrency scenarios.

[0036] Example 2: Basic Arrangement Deploy LiveFlow-IR (represented as a directed acyclic graph DAG) and load the Service Level Objective (SLO), where P95 (95th percentile latency) ≤ 1200ms, and Policy-ID points to "Live Streaming Marketing Guidelines". S2 collects and standardizes bullet comments and live chat into timestamped input objects; S3 manages and folds and labels high-frequency repetitive and uncivilized content; S4 semantic repair is performed by an intelligent agent LLM cluster to complete entity completion, dialect mapping, and paraphrasing / retelling, outputting uncertainty u and completeness i; S5 clusters within a sliding window to obtain three "standard questions" and representative statements; S6 calculates multi-objective scores and selects two direct answer paths and one clarification path based on the scores and thresholds; S9 all compliance justification vectors E reach the gating threshold; S10 binds the output with E to form justified output; S11 calculates the commitment value and writes it into the audit database, forming an irreversible trace of "problem-policy-output-justification".

[0037] Example 3: Clarification Priority and Path Shifting When S4 obtains u>0.4 or i<0.7, S6 switches the original direct answer path to a clarification path. S7 generates clarification cards containing three candidate questions (A / B / C), which are displayed at the front end and feedback is collected. Scheduling maintains a single front-end image, and the speaking order of multiple agents is controlled by the Token ring (ring speaking rights protocol). The indicator and learning module records the distribution of selections and updates the weights α / β / γ / δ online, thereby improving the direct answer hit rate of subsequent similar questions. An observation of a direct answer ratio increase of approximately 12% was made.

[0038] Example 4: Risk Budget-Driven Conservative Model Initialize the risk budget R_total=100. The concentrated occurrence of medical-related Q&A sessions accelerated the consumption of compliance risks. When R_remain<20, S12 triggers conservative mode: S6 forces high-R questions to the rejection path; S9 calls the rejection template to generate alternative suggestions and authoritative channel explanations; the SLO on the IR edge is downgraded to P95≤1500ms to ensure stable throughput; the audit module commits to recording all rejection outputs and marks them as "budget triggered".

[0039] Example 5: Speaking Right Token Ring Defense Against Concurrent Call Interception Within the same question cluster, the answering and rejecting agents complete their drafts almost simultaneously. The scheduling layer allocates speaking rights according to the Token ring: only agents holding tokens can enter the S9 compliance gate and output, while the rest wait to be released; if the timeout occurs, the Token ring reclaims the tokens and allocates them to the conservative backup agent to avoid overlapping of two sounds and semantic conflicts; after going live, subjective negative reviews about "interrupting / fighting sounds" decreased by about 80%.

[0040] Example 6: Proof-Carrying Output and Audit Commitment For each output, calculate the E vector (e.g., E={topic=0.93, sensitive=0.05, source=0.88, evidence=1}); S10 concatenates the text summary with E, timestamp, and random number, calculates the commitment H, and stores it with the session index; S11 provides an audit retrieval interface, and post-event sampling checks use H as the verification basis to verify that the output and its rationale have not been tampered with; outputs involving high-risk topics are given a higher audit priority.

[0041] Example 7: Back Pressure and Drift Control During peak sales periods, if the queue length Q > Q_max and P95 > SLO target, temporary capping is triggered in S12: the number of concurrent responses is limited and the priority of the clarification path is raised to the highest level, and non-critical issues are postponed; at the same time, the output consistency index is calculated, and if consistency anomalies (drift signals) occur N times in a row, the specific sub-model is paused and the backup agent is switched; after recovery, the capping is gradually rolled back according to the sliding window strategy.

[0042] Example 8: Cross-Session Adaptation The metrics and learning module aggregates the transaction, follow-up questions, complaints, and deletion rates from the most recent three sessions, and updates the scoring weights α / β / γ / δ and thresholds in small increments; it increases the weight ratio of A and B for "parameter comparison standard questions" and lowers the penalty coefficient of L; the direct answer hit rate improves and the violation trigger rate decreases in the next session, and the risk budget consumption is smoother.

Claims

1. A method for orchestration and compliance enforcement for live interactive, the method comprising: Comprising at least: A) constructing a live interaction intermediate representation in a directed acyclic graph, the nodes of which comprise at least collection, semantic repair, clustering, clarification, dispatch, answer, reject, and audit, and the edges of which bind service level objectives and compliance policy identifiers; B) collecting live input streams and performing governance preprocessing; C) performing semantic repair on the input by an agent large language model cluster to obtain uncertainty u and completeness i, and performing semantic clustering within a sliding window to generate canonical questions; D) computing a multi-objective score based on answerability A, business value B, compliance risk R, and latency cost L and routing to a direct answer, clarification, or reject path; E) computing a compliance evidence vector E and performing compliance gating before output, downgrading to conservative phrasing or rejection and providing alternative suggestions if threshold is not met; F) binding the output with the E vector to form an evidence-backed output, and writing a summary of "question-policy-output-evidence" to an audit footprint; G) maintaining a session risk budget and performing service level objective downgrading and priority adjustment based on budget, backpressure, and model drift detection; H) controlling the multi-agent's speech using a ring-speech protocol to ensure orderly output of a single front image.

2. The method of claim 1, wherein the multi-objective score satisfies (including but not limited to): score = α·A + β·B − γ·R − δ·L, and α, β, γ, δ are optionally updated based on online indicators.

3. The method of claim 1, wherein the compliance evidence vector E comprises at least topic consistency score, sensitive domain score, source credibility score, and evidence existence flag, and the gating criterion satisfies (including but not limited to): when w·E ≥ θ_E, allow output.

4. The method of claim 1, wherein the audit footprint uses an irreversible commitment, and the commitment value satisfies (including but not limited to): H = Hash(footprint || E || timestamp || nonce).

5. The method of claim 1, wherein the backpressure processing is triggered when at least one of any latency quantile indicator, queue length, or throughput rate meets a pre-set condition, triggering a corresponding throttling, capping, or downgrading strategy.

7. A system for implementing the method of claim 1, comprising at least:

6. The method of claim 1, wherein the semantic repair supports at least multi-lingual, multi-dialect, fragmented, broken, ambiguous, incoherent, and incomplete input completion and paraphrasing, and outputs uncertainty u and completeness i. intermediate representation constructor, intermediate representation compiler and execution engine, input collection module, governance module, semantic repair module, clustering module, scoring and routing module, clarification generation module, scheduling and speech right module, answer generation module, reject generation module, compliance gating and evidence generation module, audit and commitment module, risk budget and adaptive control module, policy library, compliance library and parameter storage module, and indicator and learning module, wherein each module is configured to perform the corresponding steps of claim 1. 8.The system of claim 7, wherein the scheduling and talk right module adopts a ring talk right protocol to serialize the talk right among multiple sub-agents, and fallback to a preset conservative output under a timeout condition. 9.A computer readable storage medium having instructions stored thereon, the instructions, when executed by a processor, causing the processor to perform the method of any one of claims 1 to 6. 10.The storage medium of claim 9, wherein the instructions further cause the processor to generate a proofed output and calculate a commitment value to bind the output with a compliance proof vector and write an audit footprint.

Images