Evidence source admission and certainty stratified information processing method and system

Abstract

This invention discloses an information processing method and system for evidence source admission and deterministic stratification. The method includes: dividing the information channel into a retrieval channel and a reasoning channel; verifying roles before assembly; rejecting non-retrieval channels using a failure-closure method; retrieving resource identifiers from the retrieval channel and writing them into the evidence source set; applying the same normalization rule to both identifiers and model reference identifiers within the set and then comparing membership relationships; based on the comparison results, convergence mode, and source credibility level, applying upper limits level by level through an ordered capping pipeline and taking the lowest value to obtain the assertion level; divergent states are not assigned levels, and illegal states are closed due to failure; the level is calculated by the server and written into the structured annotation using server-specific fields; fields with the same name in the caller's request are discarded, and the server-calculated integrity verification value is also output. This invention limits the identifiers written into the evidence source set to those actually retrieved through the retrieval channel, supporting accurate counting, recalculating levels, and traceability to specific evidence items.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and information retrieval technology, specifically to an information processing method and system for evidence source access and certainty stratification. Background Technology

[0002] In existing technologies, some approaches employ a retrieval-enhanced generation technique: first, a retrieval process obtains external data, which is then provided to a large language model to generate answers, requiring the model to include source identifiers in the answers. When multiple models produce judgments on the same proposition in parallel, a summarization process is typically used to synthesize the conclusions and provide a confidence level.

[0003] International publication number WO2025085566A1 discloses a method for generating enhanced content for legal research, whose illusion-relief function is configured to verify that information indeed comes from a certain search result. This scheme verifies the content correspondence between the model output and the search results; the source of the search results themselves is not determined by channel role access criteria, nor does it describe set membership operations after applying the same bilateral transformation to the identifiers. The disclosed claims and specification do not disclose access criteria for channel roles before evidence is pooled, do not disclose comparisons of membership relationships between citation identifiers and the evidence source set after bilateral normalization, and do not provide a deterministic hierarchical calculation mechanism that can be recalculated by established rules and is traceable.

[0004] Chinese patent CN117540811A discloses a system and method for solving the illusion problem of large language models. The approach involves first performing an initial search, generating content with source information based on the search results, then using this content as background knowledge to generate a second search term and perform an enhanced search, and finally, a correction module integrates all information for correction. Chinese patent application CN118964583A discloses a method for mitigating large model illusions, which extracts the reasoning process using thought chain cue words and combines multilingual consistency detection and word entropy uncertainty scoring to determine the illusion. However, none of the above-mentioned prior art performs admission judgment on channel roles before evidence is pooled, nor does it perform bilateral normalized comparison of membership relationships between citation identifiers and evidence source sets, and it does not provide a deterministic hierarchical calculation mechanism that can be recalculated by predetermined rules and is traceable.

[0005] The aforementioned existing technologies have at least the following drawbacks: First, they lack admission criteria for evidence sources, failing to distinguish between externally retrieved channels and models that only perform reasoning. Any source can be written into the evidence source set, making it indistinguishable in form between the resource identifiers generated by the model and the retrieved identifiers. Once entered into the set, they cannot be identified afterward. Second, they lack verification of the membership relationship between citations and the evidence source set, or they only perform text-level alignment without verifying whether the identifier is indeed in the set, resulting in a distorted count of how many sources support a statement. Third, they lack deterministic hierarchical calculation rules. The confidence level is given by the model itself, which is not recalculated and cannot be traced back to specific evidence items. Summary of the Invention

[0006] The technical problems to be solved by this invention include: how to limit the identifiers written into the evidence source set to those actually retrieved through the retrieval channel, avoiding set pollution caused by the indistinguishability between the identifiers generated by the model and the retrieved identifiers; how to ensure that the support count of assertions only includes identifiers that are indeed in the set, avoiding the mirroring of the same resource to artificially increase the support score; how to ensure that the assertion level can be recalculated by predetermined rules, with the same input always yielding the same level, and traceable to specific evidence items; how to ensure that the level calculation is not an abstract score of information content, but a constraint on data writing behavior; and how to distinguish the above constraints from existing permission isolation or tenant isolation methods.

[0007] Existing isolation methods (zero-trust gateways, access control list filtering, multi-agent sandboxes) constrain who can access which resources; this invention constrains what return results can serve as evidence, and this constraint occurs at the specific stage of assembling the evidence source set. Role-based access control is not access control based on the caller's identity, but rather on the data acquisition capability of the information channel itself; even if the caller has the highest privileges, the resource identifiers it generates through the inference channel cannot enter the evidence source set. This constraint applies to data write boundaries, not personnel access boundaries.

[0008] This invention provides an information processing method for evidence source access and certainty stratification, comprising: The information channel is divided into a retrieval channel that can perform external retrieval and a reasoning channel that only performs reasoning. Before assembling the evidence source set, the role of each information channel is checked one by one; any channel that does not belong to the retrieval channel is refused participation in the assembly and the assembly is terminated in a failure-closed manner; the termination includes discarding all evidence entries written in the assembly, not outputting the hierarchy, and returning in a predetermined failure state, without replacing any part of the result; External resource retrieval is performed through the verified retrieval channel, and the identifier of the successfully retrieved resource is written into the evidence source set; Obtain the resource identifiers referenced by the model output assertions; apply the same predetermined normalization rule to both sides of the resource identifiers within the evidence source set and the referenced resource identifiers; then perform a membership comparison; references not within the evidence source set are not counted as support; and perform deduplication on the referenced resource identifiers. For assertions that cite more than one different resource identifier, there must be at least a predetermined identifier number threshold, and the identifiers must still be within the set of evidence sources after deduplication. If the predetermined identifier number threshold is not met, the citation mode will be downgraded from multi-identifier support to single-identifier support. The server determines whether a proposition is convergent or divergent based on the structured fields output by multiple models for the same proposition according to predetermined rules. Based on the membership comparison results, the reference pattern, the convergence pattern, and the source credibility level of each resource identifier, the upper limit is applied level by level through the ordered capping pipeline and the lowest value is taken to obtain the level of the assertion; wherein, no level is assigned when the convergence pattern is divergent, and failure closure occurs when the convergence pattern is illegal; assertions that do not reach the predetermined level are not written to the evidence association storage and are not retrieved downstream; The hierarchy is calculated by the server and written into the structured annotation along with the assertion and valid resource identifier as server-specific fields. The specific fields are generated by the server during response assembly, and fields with the same name in the caller's request are discarded. The server calculates the integrity verification value of the specific fields and outputs it. The reader determines that the structured annotation has not been tampered with based on the verification value.

[0009] Optionally, the role of the information channel is predetermined by the controlled registration process; the role is stored in the read-only area of ​​the registry, and the runtime interface does not provide a method for writing to this field; any write request to this field is rejected and recorded in the audit log.

[0010] Optionally, multiple retrieval channels concurrently execute external resource retrieval, with a predetermined deadline; retrieval channels that fail to return by the deadline are marked as overdue, and the retrieved portions are merged; the return results of each retrieval channel are merged according to the predetermined channel order; when the number of successfully retrieved retrieval channels reaches the predetermined channel quantity threshold, a quantity attainment flag is set; if the predetermined channel quantity threshold is not reached, the process is not terminated, the flag is not set, and a degradation reason is recorded; the ordered capping pipeline uses the degradation reason as an input to the condition upper limit, and applies the restriction value corresponding to the degradation reason in a predetermined order when applying the condition upper limit level by level.

[0011] Optionally, the predetermined normalization rule applies the same transformation to both the resource identifiers within the evidence source set and the resource identifiers referenced by the model output assertions, so that different written forms of the same resource are classified into the same comparison form; the predetermined normalization rule includes at least one of the following options configured according to the source: unifying host case, removing default ports, removing trailing delimiters, eliminating tracking parameters, and unifying parameter order; the normalization rule retains the original identifier and records the transformation process; and specifies rejection or secondary verification for cases where collisions occur after normalization.

[0012] Optionally, the source credibility level is divided into multiple levels based on the acquisition method of the resource identifier, and the order between the levels is pre-agreed; the level is pre-bound by the controlled registration process according to the acquisition method of the channel, and the server writes it by looking up the table according to the channel identifier; only downgrading of a single result is allowed within the rules configured on the server side, and channels are not allowed to automatically upgrade; when the level declared by the channel is higher than its registered level, the registered level shall prevail; when multiple resource identifiers referenced by the same assertion belong to different source credibility levels, the strongest one in that dimension shall be taken as the input value of that dimension.

[0013] Optionally, the ordered capping pipeline receives at least the following five inputs: first, the reference pattern; second, the convergence pattern, determined by the server based on the structured fields output by multiple models according to predetermined rules; third, the number of identifiers still in the evidence source set after deduplication; fourth, the source credibility level of each identifier; and fifth, the condition upper limit corresponding to the downgrade reason. The ordered capping pipeline first synthesizes a provisional level from the reference pattern dimension, and then applies the condition upper limits of each level in a predetermined order. When applying each level, the current level is compared with the upper limit value corresponding to that level, and the lower one is taken.

[0014] Optionally, the structured annotation is exposed to the caller via a read-only interface provided by the server, which does not provide a method for writing the hierarchy fields; the calculation input for the hierarchy is taken only from the evidence entries recorded by the server, and not from the content submitted by the caller; the structured annotation also records the rule version identifier applicable to this layering, the convergence mode value, the deduplicated identifier list participating in the counting, and the application result of the upper limit at each level, so that the hierarchy can be reproduced based on the record.

[0015] Optionally, the failure closure applies to two types of non-compliance during the assembly and stratification phases. Non-compliance during the assembly phase includes the channel role not belonging to the retrieval channel and the channel not declaring a role. Non-compliance during the stratification phase includes the convergence mode being illegal and the source level value not belonging to the predetermined enumeration. Once the above situations are detected, the current assembly or stratification calculation is terminated. The termination includes discarding all evidence entries written in the current assembly, not outputting the stratification level, and returning with a predetermined failure status, without replacing any part of the result. Single-channel retrieval failure or overdue during the retrieval execution phase is not subject to failure closure, but is handled in an accounting manner and the retrieved part is merged. Insufficient support at the assertion level is handled as a downgrade and the process is not terminated.

[0016] Optionally, for assertions that do not reach the predetermined level, the resource identifier referenced by the assertion is retrieved again.

[0017] This invention also provides an information processing method for evidence source access and membership verification, comprising: The information channel is divided into a retrieval channel that can perform external retrieval and a reasoning channel that only performs reasoning. Before assembling the evidence source collection, the role of each information channel is examined one by one; any channel that does not belong to the retrieval channel is refused participation in the assembly, and the assembly is terminated in a failure-closed manner. External resource retrieval is performed through the verified retrieval channel, and the identifier of the successfully retrieved resource is written into the evidence source set; Obtain the resource identifiers referenced by the model output assertions; apply the same predetermined normalization rule to both sides of the resource identifiers within the evidence source set and the referenced resource identifiers; then perform a membership comparison; references not within the evidence source set are not counted as support; and perform deduplication on the referenced resource identifiers. For assertions that cite more than one different resource identifier, there must be at least a predetermined identifier number threshold, and the identifiers must still be within the set of evidence sources after deduplication. If the predetermined identifier number threshold is not met, the citation mode will be downgraded from multi-identifier support to single-identifier support. Based on the membership comparison results and the reference pattern, the support status of the assertion is determined and written into the structured annotation.

[0018] Optionally, the role of the information channel is predetermined by the controlled registration process; the role is stored in the read-only area of ​​the registry, and the runtime interface does not provide a method for writing to this field; any write request to this field is rejected and recorded in the audit log.

[0019] This invention also provides an information processing system for evidence source access and certainty stratification, comprising: The information channel management module is used to divide the information channel into a retrieval channel that can perform external retrieval and a reasoning channel that only performs reasoning. The evidence source access module is used to verify the role of each information channel one by one before assembling the evidence source set; any channel that does not belong to the retrieval channel is refused participation in the assembly, and the assembly is terminated in a failure closure manner. The evidence source set construction module is used to perform external resource retrieval by the retrieval channel that has passed the verification, and write the identifier of the successfully retrieved resource into the evidence source set; The membership verification module is used to obtain the resource identifiers referenced by the model output assertions. After applying the same predetermined normalization rule to both the resource identifiers in the evidence source set and the referenced resource identifiers, a membership comparison is performed. References not in the evidence source set are not counted as support. The referenced resource identifiers are deduplicated. For assertions that reference more than one different resource identifier, there must be at least a predetermined identifier number threshold that is still in the evidence source set after deduplication. If the predetermined identifier number threshold is not met, the reference mode is downgraded from multi-identifier support to single-identifier support. The deterministic layering module is used by the server to determine whether the structured fields output by the same proposition from multiple models are in a convergent or divergent state according to predetermined rules. Based on the membership comparison results, the reference pattern, the convergence pattern, and the source credibility level of each resource identifier, the upper limit is applied level by level through the ordered capping pipeline and the lowest value is taken to obtain the level of the assertion. Among them, no level is assigned when the convergence pattern is divergent, and failure is closed when the convergence pattern is illegal. Assertions that do not reach the predetermined level are not written to the evidence association storage and are not retrieved downstream. The annotation storage module is used to write the structured annotation, along with the assertion and valid resource identifier, into a server-specific field after the server calculates the level. The specific field is generated by the server during response assembly, and fields with the same name in the caller's request are discarded. The server calculates the integrity verification value of the specific field and outputs it.

[0020] This invention also provides an information processing system for evidence source access and membership verification, comprising: The information channel management module is used to divide the information channel into a retrieval channel that can perform external retrieval and a reasoning channel that only performs reasoning. The evidence source access module is used to verify the role of each information channel one by one before assembling the evidence source set; any channel that does not belong to the retrieval channel is refused participation in the assembly, and the assembly is terminated in a failure closure manner. The evidence source set construction module is used to perform external resource retrieval by the retrieval channel that has passed the verification, and write the identifier of the successfully retrieved resource into the evidence source set; The membership verification module is used to obtain the resource identifiers referenced by the model output assertions. After applying the same predetermined normalization rule to both the resource identifiers in the evidence source set and the referenced resource identifiers, a membership comparison is performed. References not in the evidence source set are not counted as support. The referenced resource identifiers are deduplicated. For assertions that reference more than one different resource identifier, there must be at least a predetermined identifier number threshold that is still in the evidence source set after deduplication. If the predetermined identifier number threshold is not met, the reference mode is downgraded from multi-identifier support to single-identifier support. The support status determination module is used to determine the support status of the assertion based on the membership comparison result and the reference pattern, and write it into the structured annotation.

[0021] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes an information processing method for evidence source access and deterministic layering, or an information processing method for evidence source access and membership verification.

[0022] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements an information processing method for evidence source access and deterministic stratification or an information processing method for evidence source access and membership verification.

[0023] The beneficial effects of this invention include: This invention performs admission judgment on the role of the information channel before assembling the evidence source set, and refuses non-retrieval channels from participating in the assembly by failure closure, so that the identifier written into the evidence source set is limited to the actual retriever via the retrieval channel. The interception occurs before the channel is executed rather than after the fact. Non-retrieval channels are not executed in the assembly interface, so there are no side effects of retrieval-type calls, no consumption of retrieval quota, and no records resembling retrieval left in the retrieval log or retrieval cache. Therefore, there is no intermediate state that requires after-the-fact verification.

[0024] This invention compares membership relationships by applying the same predetermined normalization rule to both resource identifiers within the evidence source set and identifiers referenced in the model output. This ensures that the support count of an assertion only includes identifiers that are indeed within the set, while identifiers generated by the model itself do not generate support. For assertions that reference more than one different resource identifier, there must be at least a predetermined threshold of identifiers that remain in the set after deduplication. Those that do not meet the threshold have their citation mode downgraded to the single identifier support level before entering the stratification, thereby avoiding the artificial increase in support from several mirror images of the same article and reducing the strict evidence requirements to a simple quantitative comparison.

[0025] This invention solves the technical problems in existing technologies where confidence levels cannot be recalculated or verified, the same input can yield different values ​​in different runs, and specific evidence entries cannot be traced. The assertion level is determined by the server using structured fields output from multiple models according to predetermined rules. Based on membership comparison results, citation patterns, convergence patterns, and source confidence levels, multiple inputs are processed through an ordered capping pipeline, with the lowest value applied to obtain the assertion level. This ensures that the level is recalculated according to predetermined rules, guaranteeing the same level for the same frozen input, and allowing for tracing back to specific evidence entries. No level is assigned when the convergence pattern is divergent, and failure occurs when the convergence pattern is illegal. The level is calculated by the server and written into the structured annotation, and cannot be overwritten by the caller.

[0026] This invention transforms hierarchical calculation from an abstract scoring into a technical constraint on data writing behavior by making hierarchical levels have technical consequences for system behavior. Assertions that do not reach the predetermined level are not written to the evidence association storage, are not retrieved downstream for reuse, or are triggered to be retrieved again. At the same time, through mechanisms such as server-side proprietary fields, integrity verification values, and read-only interfaces, this constraint has implementable technical means, which is different from simple permission declarations. Attached Figure Description

[0027] 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 recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of the overall process of an information processing method for evidence source access and certainty stratification provided in an embodiment of the present invention; Figure 2 A detailed flowchart illustrating the evidence source access stage provided in this embodiment of the invention; Figure 3 A detailed flowchart illustrating the membership verification stage provided in this embodiment of the invention; Figure 4 A detailed flowchart illustrating the deterministic layering stage provided in this embodiment of the invention; Figure 5 A schematic diagram of the architecture of an information processing system for evidence source access and certainty stratification provided in an embodiment of the present invention; Figure 6 This is a hardware structure diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0029] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0030] The relevant terms used in this application are explained below: An information channel refers to an input interface or data source path used to provide external resources, internal inference results, or other data sources to an information processing system. This includes external resource acquisition interfaces, data retrieval interfaces, artificial intelligence model inference interfaces, tool invocation interfaces, and other data paths capable of providing information input. Different information channels have different role attributes based on their data acquisition capabilities and intended use. Specifically, a retrieval channel is an information channel capable of acquiring data from external resource environments that can serve as sources of evidence. It performs external resource retrieval operations and returns retrieval results that include at least a resource identifier. These external resources include document resources, network resources, database records, file resources, and other identifiable data objects. An inference channel, on the other hand, is an information channel used to generate, process, or transform information inference results, but does not serve as an entry point for acquiring external evidence sources. It can output propositional assertions, inference results, citation information, and other structured or unstructured processing results. The content output by the inference channel is used for subsequent verification and hierarchical processing, but does not directly participate in the assembly of the evidence source set.

[0031] The evidence source set refers to a data set composed of resource identifiers returned through the retrieval channel after role access detection and confirmed by the system. It records data objects permitted as supporting sources in the current processing task. Its elements include at least a resource identifier, corresponding retrieval result, source level information, and other associated status information. Resources returned through information channels that have not undergone access detection are not included in this set. A resource identifier is a data identifier used to uniquely or quasi-uniquely indicate the location, identity, or association of an external resource. This includes network addresses, file identifiers, database indexes, data object numbers, and other data markers capable of locating resources. Resource identifiers are used to establish the correspondence between the model output's referenced content and the evidence source set. The referenced resource identifier is the data identifier used in the model output assertion to indicate the source on which the assertion is based. It is generated by the model output and used to verify membership relationships with the resource identifiers in the evidence source set.

[0032] Evidence source admission refers to the process of detecting whether candidate information channels are qualified to participate in evidence source construction before the evidence source set is assembled. This includes at least obtaining the information channel role, determining whether the role belongs to the retrieved channel, and allowing or denying its participation in the evidence source set assembly based on the determination result. This admission process occurs before the evidence source set is formed. Failure closure refers to a mechanism where, when a state that does not meet preset conditions is detected, the system refuses to continue executing the corresponding processing flow and does not convert the abnormal state into a valid state. It applies to illegal information channel roles, illegal convergence modes, illegal source levels, and other preset unqualified states. Failure closure differs from degradation processing; it does not produce valid results that can continue to participate in subsequent calculations.

[0033] Membership comparison refers to the process of determining whether the resource identifiers referenced in the model's output assertions belong to the evidence source set. Before performing membership comparison, the resource identifiers in the evidence source set and the referenced resource identifiers are subjected to the same normalization process to eliminate erroneous judgments caused by differences in representation. After membership comparison, references belonging to the evidence source set are counted as valid support, while references not belonging to the evidence source set are not counted in the support count. Normalization rules refer to data processing rules used to convert resource identifiers with different representations into a unified comparison form, including but not limited to standardizing host case, removing default ports, removing trailing delimiters, deleting tracking parameters, and adjusting parameter order. Resource identifiers in the evidence source set and referenced resource identifiers are processed using the same normalization rules.

[0034] Convergence mode refers to the consistency state among the outputs of multiple models for the same proposition, including at least convergent, divergent, and illegal states. A convergent state indicates that the outputs of multiple models meet a preset consistency condition; a divergent state indicates that the outputs of multiple models differ significantly; and an illegal state indicates that the output is missing or does not conform to a preset format requirement. Source credibility level refers to source level information predetermined based on the resource identifier acquisition method, source type, or source status. It is used to indicate the completeness of the evidence's source, but does not imply that the corresponding assertion itself is necessarily correct. Different resource identifiers can have different source credibility levels.

[0035] Deterministic stratification refers to the process of generating preset stratification labels for model output assertions based on the state of evidence support. This is based at least on membership comparison results, the number of valid resource labels, multi-model convergence modes, and source credibility levels. The deterministic stratification results are calculated by system rules, not declared by the model itself. Ordered capping pipeline refers to the process of applying constraints to the initial stratification level in a predetermined order, selecting results not exceeding the current stratification level as the new stratification state at each constraint level. This includes inputting multiple stratification dimensions, generating the initial stratification level, applying constraints sequentially, and outputting the final stratification level. The order in which different constraints are applied is determined by system preset rules.

[0036] Structured annotations refer to data structures generated by the server based on deterministic layering results, used to describe assertion levels and associated evidence information. These annotations must at least include the corresponding assertion, valid resource identifiers, source credibility levels, and layering results. The structured annotations are generated by the server, and the caller only has read permissions. The number of valid supporting sources refers to the number of resource identifiers actually existing in the evidence source set after resource identifier extraction, normalization, membership comparison, and deduplication. References without membership confirmation are not included in the number of valid supporting sources.

[0037] like Figure 1 As shown, this embodiment provides an information processing method for evidence source admission and certainty layering. The method is divided into three sequentially connected processing stages, corresponding to evidence source admission, verification of citations and set membership relationships, and certainty credibility layering. The outputs of the three stages serve as the inputs to the next stage, and are coupled along the same data integrity link.

[0038] like Figure 2 As shown, during the evidence source access phase, the system first receives the proposition to be processed and multiple information channels. Information channels are divided into two roles within the system: retrieval channels, which can perform external retrieval and return the retrieved content and its resource identifier; and inference channels, which only perform inference based on their own parameters and do not perform external retrieval. The role of a channel is predetermined by the controlled registration process and provided to the system in read-only mode. A fixed role type is assigned to a channel during registration, and neither the caller nor the information channel itself has the right to modify this role type during runtime. The role is stored in the read-only area of ​​the registry; the runtime interface does not provide a write method for this field; any write request to this field is rejected and recorded in the audit log.

[0039] Before assembling the evidence source set, the system verifies the role of each information channel entering the assembly process. Any channel whose role does not belong to the retrieval channel, including those without declared roles, is rejected before any channel is executed, and the assembly is terminated with a failure closure. This failure closure is not a downgraded release, but a direct termination of the assembly process, preventing the results of unqualified channels from being included in the evidence source set. Termination includes discarding all evidence entries already written in the assembly, not outputting a hierarchy, and returning with a predetermined failure status, without replacing any partial results. The above role-based access control only constrains the set of input channels for the evidence source set assembly interface. Inference channels execute in a separate inference interface, and their outputs are entered into subsequent processes as assertions and references to be verified. They do not participate in the writing of the evidence source set and are therefore not rejected due to role-based access control.

[0040] The retrieval channel that passes role verification performs the retrieval of external resources. Multiple retrieval channels can execute retrieval operations concurrently, each with a predetermined deadline. Retrieval channels that fail to return by the deadline are marked as overdue, and the retrieved portions are merged. The return results from each retrieval channel are merged according to the predetermined channel order, rather than the actual return time order of each channel, thus making the assembly result independent of the completion order of each channel.

[0041] When the number of successful retrieval channels reaches a predetermined threshold, the system sets a threshold-reached flag. If the threshold is not reached, the system does not terminate the process but does not set the flag and records a degradation reason. This degradation reason is used to limit the level to which assertions can be assigned in subsequent deterministic layering. A resource identifier is counted as an evidence entry and written into the evidence source set only when it is returned by a retrieval channel in a successful state and is returned as part of the retrieval result along with the retrieval content summary. Therefore, identifiers written into the evidence source set are limited to those successfully retrieved via a retrieval channel and returned along with the retrieval result.

[0042] like Figure 3 As shown, during the membership verification phase, the system acquires assertion information from multiple models and the corresponding referenced resource identifiers. The same predefined normalization rule is applied to both the resource identifiers in the evidence source set and the referenced resource identifiers in the assertion information to obtain a normalized form for comparison. This predefined normalization rule includes at least one of the following options configured by source: standardizing host case sensitivity, removing default ports, removing trailing delimiters, eliminating tracking parameters, and standardizing parameter order. The normalization rule retains the original identifiers and records the transformation process; it specifies rejection or secondary verification for collisions occurring after normalization. Specific items of the normalization rule can be configured one by one according to the source.

[0043] After normalization, the system performs a membership comparison to determine whether the normalized cited resource identifiers belong to the evidence source set. Citations not in the set are not counted in the number of valid supporting sources for the assertion. The system deduplicates the cited resource identifiers. For assertions citing more than one different resource identifier, there must be at least a predetermined identifier count threshold that remains within the evidence source set after deduplication. Those not meeting this threshold have their citation mode downgraded from multi-identifier support to single-identifier support. This lower level then serves as input for the deterministic stratification stage, participating in ordered capping to arrive at a lower level. This quantity threshold does not apply to citations supported only by a single identifier.

[0044] like Figure 4 As shown, in the deterministic layering stage, the system receives multiple inputs. The first is the reference mode, derived from the aforementioned membership verification and quantity threshold determination. The second is the convergence mode, determined by the server based on the structured fields output by multiple models for the same proposition, according to predetermined rules, to determine whether it is a convergent or divergent state. The server does not accept the semantic correctness self-reported by the models; any missing, empty, or non-predetermined values ​​are considered illegal. The third is the number of identifiers referenced by the assertion and still within the evidence source set after deduplication. The fourth is the source credibility level of each identifier, which is divided into multiple levels based on the acquisition method of the resource identifier, with the order between levels pre-defined. These levels are pre-bound by the controlled registration process according to the channel acquisition method, and the server writes them by looking up the table based on the channel identifier; only downgrading of a single result is allowed within the rules configured on the server side, and channels are not allowed to automatically upgrade; when the declared level of a channel is higher than its registered level, the registered level prevails. When multiple resource identifiers referenced by the same assertion belong to different source credibility levels, the strongest one within that dimension is taken as the input value for that dimension. Fifthly, there is the upper limit of the conditions corresponding to the reason for downgrading.

[0045] The system derives the assertion hierarchy through an ordered capping pipeline. This pipeline first synthesizes a provisional hierarchy from dimensions such as reference patterns, and then applies conditional upper limits level by level in a predetermined order. At each level, the current level is compared with its corresponding limit value, and the lower value is selected. The triggering conditions and corresponding upper limit values ​​for each level are determined according to a predetermined truth table. When the provisional hierarchy is no higher than level N, the upper limit at level M is not applied. Therefore, the order in which each level is applied affects the final result; this is the meaning of "ordered." This pipeline is not simply a matter of taking the minimum value of each upper limit.

[0046] When the convergence mode is divergent, the system does not assign a hierarchy to assertions, but instead labels them with information content categories. When the convergence mode is illegal, the system performs a failure closure, neither assigning a hierarchy nor granting an exemption; the termination includes discarding all evidence entries written in this assembly, not outputting a hierarchy, and returning with a predetermined failure status, without replacing any partial result.

[0047] The hierarchy is calculated by the server and written into the structured annotation. After calculation, the hierarchy, along with the assertions and valid resource identifiers, is written into the structured annotation as a server-specific field. This specific field is generated by the server during response assembly; fields with the same name in the caller's request are discarded. The server calculates an integrity check value for each specific field and outputs it. The reader uses this check value to determine if the structured annotation has been tampered with. This structured annotation is associated with the corresponding assertion, valid resource identifier, and source credibility level. The structured annotation is exposed to the caller through a read-only interface provided by the server; this interface does not provide a method for writing the hierarchy field. The input for hierarchy calculation is only taken from the evidence entries recorded by the server, not from the content submitted by the caller.

[0048] The structured annotation also records the rule version identifier applicable to this stratification, the convergence mode value, the deduplicated identifier list participating in the counting, and the application results of the upper limits at each level, so that the stratification can be reproduced based on the record. This record is stored together with the assertion, valid resource identifier, and source credibility level, forming a complete audit chain traceable to specific evidence items.

[0049] Assertions that do not reach the predetermined level are not written to the evidence association storage, are not retrieved downstream for reuse, or trigger the re-retrieval of the resource identifier referenced by the assertion. Therefore, hierarchical calculation is not an abstract score of information content, but a technical constraint on data writing behavior.

[0050] Failure closure applies to two types of non-compliance during the assembly and stratification phases. Non-compliance during the assembly phase includes situations where the channel role does not belong to the retrieval channel and the channel has not declared a role. Non-compliance during the stratification phase includes situations where the convergence mode is illegal and the source level value does not belong to the predetermined enumeration. Once any of these situations is detected, the current assembly or stratification calculation is terminated. Termination includes discarding all evidence entries written in the current assembly, not outputting the stratification level, and returning with a predetermined failure status, without replacing any partial result. Single-channel retrieval failures or expirations during the retrieval execution phase are not subject to failure closure; instead, they are handled according to the accounting method, and the retrieved portions are merged. Insufficient support at the assertion level is handled according to the aforementioned downgrade process, and the process is not terminated.

[0051] The aforementioned evidence source access, membership verification, and deterministic layering are functionally coupled along the same data integrity chain. The layering identifier input can only be generated through membership comparison and deduplication; the source level input can only be assigned by the write stage constrained by role access; the assertion corresponding to the convergence mode input must be verified by membership before entering the layering process. Failure closure of role access constitutes the operational prerequisite for the latter two. These three are not assembled from independent modules, but rather are interdependent along the same chain. The execution location and constrained objects of each step in the system include: intercepting constraint write behavior, rejecting constraint system state transitions, and defining the writable boundaries of hierarchical write constraint data.

[0052] Example of server-specific fields and integrity verification: The hierarchy is calculated by the server and written into the structured annotation along with the assertion and valid resource identifier as a server-specific field. This specific field is generated by the server during response assembly, and its field name uses the server's internal namespace prefix (e.g., "svc:level"). Fields with the same name in the caller's request are discarded during the assembly stage. The server calculates an integrity verification value (e.g., HMAC-SHA256) for this specific field and outputs it. The reader uses this verification value to determine that the structured annotation has not been tampered with during transmission and storage. The structured annotation is stored in a read-only storage area controlled by the server. The caller obtains it through a read-only interface. This interface does not provide a method for writing the hierarchy field; any write request is rejected by the server and an error status is returned.

[0053] The reproducible and traceable technical carrier: The structured annotation also records the rule version identifier (e.g., "rule-v3.2") applicable to this stratification, the convergence mode value, the deduplicated identifier list participating in the counting, and the application results of the upper limits at each level, enabling the stratification to be reproduced based on the record. This record, along with assertions, valid resource identifiers, and source credibility levels, is stored together to form a complete audit chain traceable to specific evidence items.

[0054] Specific example of an ordered capping pipeline: Assume the levels are divided into four grades: A, B, C, and D, from highest to lowest, with the order A>B>C>D>skip. Reference pattern dimension: Multiple identifiers supporting the pipeline and reaching a certain quantity threshold are combined to form a provisional grade A; multiple identifiers supporting the pipeline but not reaching the quantity threshold are downgraded to grade B; single identifier support is combined to form a provisional grade C. Convergence pattern dimension: No additional restrictions are imposed in the convergent state; in the divergent state, the output is directly output without being assigned a grade. Source credibility level dimension: The upper limit for first-level sources is A, the upper limit for second-level sources is B, and the upper limit for third-level sources is C. Degradation reason dimension: When a degradation reason exists, the upper limit B is applied. The capping pipeline is applied level by level in the following order: Level 1 is reference pattern synthesis; Level 2 is the upper limit of source credibility level; Level 3 is the upper limit of degradation reason; Level 4 is convergence pattern determination (covering all of the above in the divergent state). When applying each level, the lower of the current grade and the upper limit is taken. If the provisional level is already C, then the upper limit of the downgrade reason B will no longer take effect (because B is higher than C, the lower one will still be C).

[0055] Example 1 (Divergent Path): The provisional level is synthesized into A using the reference pattern, with a source credibility level of Level 1 (upper limit A) and no reason for downgrading. The fourth level determines the convergence pattern as divergent, and the output is directly output without assigning a level.

[0056] Example 2 (Path with Insufficient Quantity): The provisional level is synthesized into A (multi-identifier support) using the reference pattern, with a source credibility level of Level 1 (upper limit A), and a downgrade reason exists (upper limit B). After applying Level 3, the level drops to B. The convergence mode is in a convergent state, and the final output level is B.

[0057] In one improvement process, the early design weighted the number of search results, but it was found that the support score could be artificially inflated by mirroring and spam results, so it was changed to stratification based on source reliability. Initially, a single model provided the confidence score, but it was found that the model could easily endorse itself, so it was changed to apply consistency constraints on the convergence patterns of multiple independent models on the same proposition. Initially, only the model was required to mark the source of the citation and verify the content summary, but it was found that the model could generate identifiers that looked like real links and were indistinguishable from real retrieval identifiers in form, so it was changed to conduct admission judgment on the channel role before the evidence was entered into the pool, and after applying normalization rules to both the identifiers in the set and the citation identifiers, the membership relationship was compared, and citations not in the set were not counted as support.

[0058] Several components of this invention can be replaced by equivalent means, specifically as follows: the role criterion can be declarative or capability-verified; the merging order can be ordered according to a predetermined channel order or according to the resource identifier content summary; the execution method can be multi-channel concurrent or sequential execution; the quantity attainment judgment can be equal-weighted counting or weighted counting; the normalization rule can add or remove specific items or configure different whitelists for each source; the comparison key can be the resource identifier or a joint key of the resource identifier and the retrieved content summary; the resource identifier form can be a Uniform Resource Locator, a Numerical Object Identifier, or a Content Addressing Hash; the missing reference handling can downgrade references not in the set to a lower level or exclude them completely; the capping rule can be to take the lowest value at each level, weighted summation and then classify by threshold or take the median; the input dimension can be increased by the freshness of the source publication time; the divergent state handling can be changed to labeling by information volume category or directly assigning the lowest level; the level order can be a predetermined fixed order or a configurable order; the content summary algorithm can be any algorithm with anti-collision properties; the calculation subject can be undertaken by the server or an independent process or a trusted execution environment.

[0059] like Figure 5 As shown, this embodiment provides an information processing system for evidence source access and deterministic stratification, including an information channel management module, an evidence source access module, an evidence source set construction module, a member relationship verification module, a deterministic stratification module, and an annotation storage module.

[0060] The information channel management module divides information channels into retrieval channels capable of external resource retrieval and inference channels that only perform inference. The evidence source admission module verifies the role of each information channel before assembling the evidence source set, rejecting non-retrieval channels from participating in the assembly and terminating the assembly process with a failure-closed approach. This termination includes discarding all evidence entries already written in the assembly, not outputting a hierarchy, and returning with a predetermined failure status, without replacing any partial result. The evidence source set construction module allows retrieval channels that pass the verification to perform external resource retrieval and writes the identifier of successfully retrieved resources into the evidence source set.

[0061] The membership verification module is used to obtain the resource identifiers referenced by the model output assertions, perform normalization processing on the resource identifiers in the evidence source set and the resource identifiers referenced by the assertions, and determine the membership relationship between them; references not in the evidence source set are not counted as support; and deduplication processing is performed on the referenced resource identifiers. The deterministic layering module is used by the server to determine whether the structured fields output by multiple models for the same proposition are in a convergent or divergent state according to predetermined rules; based on the membership comparison results, reference patterns, model convergence patterns, number of valid resource identifiers, and source credibility levels, the assertion level is calculated using an ordered capping method; when the convergence pattern is divergent, no level is assigned; when the convergence pattern is illegal, failure closure occurs; assertions that do not reach the predetermined level are not written to the evidence association storage, are not retrieved downstream for reuse, or trigger the re-retrieval of the resource identifiers referenced by the assertion. The annotation storage module is used for the server to calculate the hierarchy and write it into structured annotations, and restricts the caller from modifying the annotations at this level; after the hierarchy is calculated by the server, it is written into the structured annotations along with the assertion and valid resource identifier as server-specific fields; the specific fields are generated by the server during response assembly, and fields with the same name in the caller's request are discarded; the server calculates the integrity verification value of the specific fields and outputs it together.

[0062] Each module of the system executes the corresponding steps in the above method embodiments, and its implementation principle and technical effect will not be elaborated here.

[0063] like Figure 6 As shown, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned information processing method for evidence source access and deterministic layering, or the information processing method for evidence source access and membership verification. The electronic device may include at least one processor, a memory, a RAM, and a communication interface, and at least one processor, a memory, a RAM, and a communication interface are connected together via a bus. At least one processor executes at least one computer-readable instruction stored or encoded in the memory.

[0064] It should be understood that the computer-executable instructions stored in memory, when executed, cause at least one processor to perform the above-described combinations in the various embodiments of this specification. Figures 1 to 5 The description includes various operations and functions.

[0065] In the embodiments of this specification, electronic devices may include, but are not limited to: personal computers, server computers, workstations, desktop computers, laptop computers, notebook computers, mobile electronic devices, smartphones, tablet computers, cellular phones, personal digital assistants, handheld systems, messaging devices, wearable electronic devices, consumer electronic devices, etc.

[0066] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-mentioned information processing method for evidence source access and deterministic stratification, or information processing method for evidence source access and membership verification.

[0067] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0068] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An information processing method for evidence source access and certainty stratification, characterized in that, include: The information channel is divided into a retrieval channel that can perform external retrieval and a reasoning channel that only performs reasoning. Before assembling the evidence source set, the role of each information channel is checked one by one; any channel that does not belong to the retrieval channel is refused participation in the assembly and the assembly is terminated in a failure-closed manner; the termination includes discarding all evidence entries written in the assembly, not outputting the hierarchy, and returning in a predetermined failure state, without replacing any part of the result; External resource retrieval is performed through the verified retrieval channel, and the identifier of the successfully retrieved resource is written into the evidence source set; Obtain the resource identifiers referenced by the model output assertions; apply the same predetermined normalization rule to both sides of the resource identifiers within the evidence source set and the referenced resource identifiers; then perform a membership comparison; references not within the evidence source set are not counted as support; and perform deduplication on the referenced resource identifiers. For an assertion that cites more than one different resource identifier, there must be at least a predetermined number of identifiers that remain in the set of evidence sources after deduplication; For those whose number of identifiers does not reach the predetermined threshold, their referencing mode will be downgraded from multi-identifier support to single-identifier support. The server determines whether a proposition is convergent or divergent based on the structured fields output by multiple models for the same proposition according to predetermined rules. Based on the membership comparison results, the reference pattern, the convergence pattern, and the source credibility level of each resource identifier, the upper limit is applied level by level through the ordered capping pipeline and the lowest value is taken to obtain the level of the assertion; wherein, no level is assigned when the convergence pattern is divergent, and failure closure occurs when the convergence pattern is illegal; assertions that do not reach the predetermined level are not written to the evidence association storage and are not retrieved downstream; The hierarchy is calculated by the server and written into the structured annotation along with the assertion and valid resource identifier as server-specific fields. The proprietary fields are generated by the server during response assembly, and fields with the same name in the caller's request are discarded. The server calculates an integrity check value for the proprietary fields and outputs it. The reader uses this check value to determine that the structured annotation has not been tampered with.

2. The information processing method for evidence source access and certainty stratification according to claim 1, characterized in that, The role of the information channel is predetermined by the controlled registration process; the role is stored in the read-only area of ​​the registry, and the runtime interface does not provide a method for writing to this field; any write request to this field will be rejected and recorded in the audit log.

3. The information processing method for evidence source access and certainty stratification according to claim 1, characterized in that, Multiple retrieval channels concurrently execute the retrieval of external resources, with a predetermined deadline; retrieval channels that fail to return by the deadline are marked as overdue, and the retrieved portions are merged; the return results of each retrieval channel are merged according to the predetermined channel order; When the number of successfully retrieved channels reaches the predetermined channel quantity threshold, the quantity threshold is set; if the predetermined channel quantity threshold is not reached, the process is not terminated, the flag is not set, and the degradation reason is recorded; the ordered capping pipeline uses the degradation reason as an input of the condition upper limit, and applies the limit value corresponding to the degradation reason in a predetermined order when applying the condition upper limit level by level.

4. The information processing method for evidence source access and certainty stratification according to claim 1, characterized in that, The predefined normalization rules apply the same transformation to both the resource identifiers within the evidence source set and the resource identifiers referenced by the model output assertions, so that different written forms of the same resource are classified into the same comparison form; the predefined normalization rules include unifying host case and removing default ports; The normalization rule includes at least one of the following options configured by source: removing tail delimiters, eliminating tracking parameters, and uniformly processing the parameter order; the normalization rule retains the original identifier and records the transformation process; and it specifies rejection or secondary verification for cases where collisions occur after normalization.

5. The information processing method for evidence source access and certainty stratification according to claim 1, characterized in that, The source credibility level is divided into multiple levels based on the acquisition method of the resource identifier, and the order between the levels is pre-agreed. The level is pre-bound by the controlled registration process according to the acquisition method of the channel, and the server writes it by looking up the table based on the channel identifier. Only downgrading of a single result is allowed within the rules configured on the server side, and channels are not allowed to upgrade on their own. When the level declared by a channel is higher than its registered level, the registered level shall prevail. When multiple resource identifiers referenced by the same assertion belong to different source credibility levels, the strongest one in that dimension shall be taken as the input value of that dimension.

6. The information processing method for evidence source access and certainty stratification according to claim 1, characterized in that, The ordered capping pipeline receiving includes at least the following five inputs: first, the reference mode; second, the convergence mode, which is determined by the server based on the structured fields output by multiple models according to predetermined rules; third, the number of identifiers that are still in the evidence source set after deduplication; and fourth, the source credibility level of each identifier. Fifthly, the upper limit of the conditions corresponding to the reasons for downgrading; The ordered capping pipeline first synthesizes a provisional level from the reference mode dimension, and then applies the upper limit of each level of conditions in a predetermined order. When applying each level, the current level is compared with the upper limit value corresponding to that level and the lower one is taken.

7. The information processing method for evidence source access and certainty stratification according to claim 1, characterized in that, The structured annotations are exposed to the caller via a read-only interface provided by the server. This interface does not provide a method for writing hierarchical fields. The input for hierarchical calculations is taken only from the evidence entries recorded by the server, and not from the content submitted by the caller. The structured annotation also records the rule version identifier applicable to this layering, the convergence mode value, the deduplicated identifier list participating in the counting, and the application result of the upper limit at each level, so that the layer can be reproduced based on the record.

8. The information processing method for evidence source access and certainty stratification according to claim 1, characterized in that, The failure closure applies to two types of non-compliance during the assembly and stratification phases. Non-compliance during the assembly phase includes the channel role not belonging to the retrieval channel and the channel not declaring a role. Non-compliance during the stratification phase includes the convergence mode being illegal and the source level value not belonging to the predetermined enumeration. Once the above situations are detected, the current assembly or stratification calculation will be terminated. The termination includes discarding all evidence entries written in the current assembly, not outputting the stratification level, and returning with a predetermined failure status, without replacing any part of the result. If a single-channel retrieval fails or expires during the retrieval execution phase, failure closure is not applied; instead, it is handled through accounting and the retrieved portion is merged. If there is insufficient support at the assertion level, it is handled as a downgrade and the process is not terminated.

9. The information processing method for evidence source access and certainty stratification according to claim 1, characterized in that, For assertions that do not reach the predetermined level, the resource identifier referenced by the assertion is retrieved again.

10. An information processing method for evidence source access and membership verification, characterized in that, include: The information channel is divided into a retrieval channel that can perform external retrieval and a reasoning channel that only performs reasoning. Before assembling the evidence source collection, the role of each information channel is examined one by one; any channel that does not belong to the retrieval channel is refused participation in the assembly, and the assembly is terminated in a failure-closed manner. External resource retrieval is performed through the verified retrieval channel, and the identifier of the successfully retrieved resource is written into the evidence source set; Obtain the resource identifiers referenced by the model output assertions; apply the same predetermined normalization rule to both sides of the resource identifiers within the evidence source set and the referenced resource identifiers; then perform a membership comparison; references not within the evidence source set are not counted as support; and perform deduplication on the referenced resource identifiers. For assertions that cite more than one different resource identifier, there must be at least a predetermined identifier number threshold, and the identifiers must still be within the set of evidence sources after deduplication. If the predetermined identifier number threshold is not met, the citation mode will be downgraded from multi-identifier support to single-identifier support. Based on the membership comparison results and the reference pattern, the support status of the assertion is determined and written into the structured annotation.

11. The information processing method for evidence source access and membership verification according to claim 10, characterized in that, The role of the information channel is predetermined by the controlled registration process; the role is stored in the read-only area of ​​the registry, and the runtime interface does not provide a method for writing to this field; any write request to this field will be rejected and recorded in the audit log.

12. An information processing system for evidence source access and certainty stratification, characterized in that, include: The information channel management module is used to divide the information channel into a retrieval channel that can perform external retrieval and a reasoning channel that only performs reasoning. The evidence source access module is used to verify the role of each information channel one by one before assembling the evidence source set; any channel that does not belong to the retrieval channel is refused participation in the assembly, and the assembly is terminated in a failure closure manner. The evidence source set construction module is used to perform external resource retrieval by the retrieval channel that has passed the verification, and write the identifier of the successfully retrieved resource into the evidence source set; The membership verification module is used to obtain the resource identifiers referenced by the model output assertions, apply the same predetermined normalization rule to both sides of the resource identifiers in the evidence source set and the referenced resource identifiers, and then perform membership comparison; references not in the evidence source set are not counted as support; deduplication is performed on the referenced resource identifiers; for assertions that reference more than one different resource identifier, there must be at least a predetermined identifier number threshold, and the deduplicated identifiers must still be in the evidence source set; For those whose number of identifiers does not reach the predetermined threshold, their referencing mode will be downgraded from multi-identifier support to single-identifier support. The deterministic layering module is used by the server to determine whether the structured fields output by the same proposition from multiple models are in a convergent or divergent state according to predetermined rules. Based on the membership comparison results, the reference pattern, the convergence pattern, and the source credibility level of each resource identifier, the upper limit is applied level by level through the ordered capping pipeline and the lowest value is taken to obtain the level of the assertion. Among them, no level is assigned when the convergence pattern is divergent, and failure is closed when the convergence pattern is illegal. Assertions that do not reach the predetermined level are not written to the evidence association storage and are not retrieved downstream. The annotation storage module is used to write the structured annotation, along with the assertion and valid resource identifier, into a server-specific field after the server calculates the level. The specific field is generated by the server during response assembly, and fields with the same name in the caller's request are discarded. The server calculates the integrity verification value of the specific field and outputs it.

13. An information processing system for evidence source access and membership verification, characterized in that, include: The information channel management module is used to divide the information channel into a retrieval channel that can perform external retrieval and a reasoning channel that only performs reasoning. The evidence source access module is used to verify the role of each information channel one by one before assembling the evidence source set; any channel that does not belong to the retrieval channel is refused participation in the assembly, and the assembly is terminated in a failure closure manner. The evidence source set construction module is used to perform external resource retrieval by the retrieval channel that has passed the verification, and write the identifier of the successfully retrieved resource into the evidence source set; The membership verification module is used to obtain the resource identifiers referenced by the model output assertions, apply the same predetermined normalization rule to both sides of the resource identifiers in the evidence source set and the referenced resource identifiers, and then perform membership comparison; references not in the evidence source set are not counted as support; deduplication is performed on the referenced resource identifiers; for assertions that reference more than one different resource identifier, there must be at least a predetermined identifier number threshold, and the deduplicated identifiers must still be in the evidence source set; For those whose number of identifiers does not reach the predetermined threshold, their referencing mode will be downgraded from multi-identifier support to single-identifier support. The support status determination module is used to determine the support status of the assertion based on the membership comparison result and the reference pattern, and write it into the structured annotation.

14. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the information processing method for evidence source access and deterministic stratification as described in any one of claims 1 to 8, or the information processing method for evidence source access and membership verification as described in any one of claims 10 to 11.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the information processing method for evidence source access and deterministic stratification as described in any one of claims 1 to 8, or the information processing method for evidence source access and membership verification as described in any one of claims 10 to 11.

Images