Automated Revenue Settlement System for Open-Source License-Based Expert Module Weights

Abstract

The present invention comprises: a weight public register in which an expert module developer registers MoE expert module weights independently trained by himself in an on-chain registry along with open-source license conditions and cryptographic hash values; an automatic license condition detection module that automatically detects and records license conditions at the time when a MoE model operator downloads the weights and integrates them into an expert pool; an on-chain settlement smart contract that automatically remits inference service fees to the original developer in proportion to the frequency of selection of the expert module's gating router; and a contribution measurement oracle that guarantees the integrity of the gating log using zero-knowledge proofs and commit-rebuild methods. According to the present invention, an AI knowledge economy infrastructure is realized in which even small-scale AI developers can disclose their expertise in the form of weights and automatically and continuously receive a portion of the inference service revenue.

Description

Technology Field

[0001] The present invention relates to a system that publicly registers the weights of an expert module, which is a component of an artificial intelligence (AI) model, along with open source license conditions, and automatically settles the revenue generated when said weights are integrated into a third-party Mixture of Experts (MoE) model and utilized in inference services to the original developer. More specifically, the present invention relates to a distributed ledger-based AI knowledge economy infrastructure that implements a transparent and automated reward system for the intellectual contributions of an expert module developer (300) by combining a blockchain-based on-chain registry (150), a smart contract (130), and a Zero-Knowledge Proof-based contribution measurement oracle (140). Background Technology

[0002] Mixed Expert (MoE) architecture is a technology that achieves high model performance at an inference cost not proportional to the total number of parameters by deploying multiple expert networks within a single large language model and dynamically selecting and activating the most suitable expert module through a gating router (181) based on the characteristics of the input token. This architecture has become a de facto standard structure as it is adopted by modern large language models such as Mixtral and DeepSeek-V3, and in MoE models, each expert module contributes decisively to the overall performance of the model as a set of parameters specialized for a specific domain or language pattern.

[0003] However, in the current AI development ecosystem, the dominant structure involves large AI companies training all expert module weights in-house. Consequently, even if small research teams or individual developers train exceptional expert modules in specific domains, there is no avenue for them to contribute these modules to large AI operators' MoE models and generate sustainable revenue. A widespread problem exists where researchers who have accumulated years of data in highly specialized domains—such as medicine, law, and semiconductor design—are unable to commercially realize the value of their expertise, even after implementing it in the form of AI weights, due to a lack of technical and institutional foundations.

[0004] In the field of open-source software, systems for automatically enforcing conditions regarding the use, distribution, and modification of source code through licenses such as GPL, MIT, and Apache have been in operation for decades. However, unlike source code, the legal status of AI weights is unclear; once a weight file is downloaded and physically integrated into another user's model, it is difficult to track the fact of its use, and even if tracked, there is no objective means to measure how much the weight actually contributed to inference. Existing technologies, such as blockchain-based IP registration systems, focus on registering ownership of creative works and distributing royalties at the time of selling derivative works; consequently, they fail to provide a mechanism to automatically detect when weights are integrated into the Expert pool of a MoE model and to continuously distribute inference service fees in proportion to contribution.

[0005] Against this backdrop, there is an urgent need for a new system that applies an automatic enforcement mechanism of open source licenses to AI weights, generates an unmanipulable settlement basis through a zero-knowledge proof-based contribution measurement oracle (140), and automatically distributes profits through a smart contract (130). The problem to be solved

[0006] The present invention is devised to solve the above-mentioned problems and aims to provide a system in which a weight trained by an expert module developer (300) is registered in an on-chain registry (150) along with open source license conditions, and when a MoE model operator (400) downloads the weight and integrates it into their expert pool (183), the license conditions are automatically detected and recorded, and subsequently, the inference service usage fee is automatically settled to the original developer in proportion to the selection frequency of the gating router (181).

[0007] Furthermore, the present invention aims to establish an infrastructure that enables the substantial enforcement of license obligations without voluntary participation by providing means to automatically detect on-chain unauthorized use by MoE operators integrating external weights while in an unlicensed state and to restrict access.

[0008] Furthermore, the present invention aims to generate a reliable settlement basis that cryptographically guarantees the integrity of contribution measurement data by combining Zero-Knowledge Proof technology with a Commit-Reveal method to prevent manipulation of gating logs by the operator. means of solving the problem

[0009] The open source Expert license automatic settlement system (100) of the present invention for achieving the above purpose includes a weight public register (110), a license condition automatic detection module (120), an on-chain settlement smart contract (130), and a contribution measurement oracle (140).

[0010] The weight public register (110) allows an expert module developer (300) to register their expert module weights in a public repository along with open source license conditions including revenue sharing obligations for commercial use, and records the cryptographic hash of the weights and the developer's wallet address in the on-chain registry (150).

[0011] The license condition automatic detection module (120) queries the on-chain registry (150) at the time when the MoE model operator (400) downloads expert module weights from a public repository and integrates them into the expert pool of his MoE model (180), automatically verifies the license conditions and revenue sharing ratio of the weights, and records the fact of integration. Additionally, the license condition automatic detection module (120) includes an unauthorized integration detection function to detect weights integrated without registered license conditions and records them on-chain.

[0012] The on-chain settlement smart contract (130) automatically transfers an amount corresponding to the distribution ratio specified in the license conditions from the usage fee generated in the inference service of the MoE model (180) to the original developer's wallet address in proportion to the actual inference contribution of the expert module.

[0013] The contribution measurement oracle (140) records the selection frequency of the gating router (181) per inference request for each of the multiple expert modules integrated into the MoE model (180) and verifies this with a zero-knowledge proof to use as a weight for profit distribution. Effects of the invention

[0014] According to the present invention, an expert module developer establishes a sustainable revenue generation structure in which revenue is automatically attributed regardless of which MoE model the weight is integrated into and how much inference it contributes, simply by registering the weights they have trained along with open source license conditions. This provides a strong economic incentive for AI researchers and domain experts to train and disclose specialized weights, and has the effect of dramatically improving the quality and diversity of the open source AI ecosystem.

[0015] Furthermore, by enforcing open-source license terms as code, this invention implements a fully automated rights enforcement system without the legal disputes, manual contract execution, or intervention of intermediary settlement agencies that were inevitable in conventional AI model weight trading. As a result, expert module developers can protect their rights without legal costs, and small research teams or individual developers can secure revenue on a legal basis equivalent to that of large AI operators.

[0016] Furthermore, since manipulation of the gating log by the operator side becomes cryptographically impossible through the zero-knowledge proof-based contribution measurement oracle (140) and the commit-rebuild batch aggregator (142), the reliability of the settlement basis is mathematically guaranteed without third-party verification. This has the effect of preventing settlement disputes at the source and increasing the trust of ecosystem participants.

[0017] In addition, through the combination of the unauthorized integration detector (160) and the access blocking mechanism (162), reputation-based sanctions are automatically applied to operators who do not comply with license obligations, thereby restricting their access to other Expert weights within the ecosystem. This realizes a self-enforcing governance structure that induces license compliance without the need for separate judicial enforcement measures.

[0018] Furthermore, the progressive revenue distribution structure (Clause 6) naturally incentivizes developers to train high-quality weights frequently used in actual inference by guaranteeing a higher return rate for high-contribution expert modules. Additionally, the automated revenue reinvestment feature (Clause 9) immediately converts a portion of settlement revenue into new training computing credits, forming a virtuous cycle of innovation where developers iteratively improve their experts using revenue as a resource. Brief explanation of the drawing

[0019] Fig. 1 It is a block diagram representing the overall system configuration, showing the total data flow between the four major components and external actors. Fig. 2 This is a flowchart of the weight public registration procedure, showing the sequential flow from the developer terminal (300) to the completion of registration in the on-chain registry (150). Fig. 3 The on-chain registry (150) data structure diagram shows the entire field configuration of the Expert unit. Fig. 4 This is an operation flowchart of the automatic license condition detection module (120), showing a branching structure from download to settlement smart contract activation. Fig. 5 This is a diagram showing the coexistence structure of the internal expert module pool (182) and the external expert module pool (183) within the MoE model (180) and the dynamic routing flow of the gating router (181). Fig. 6 This is an operational structure diagram of the contribution measurement oracle (140), showing the flow from the lightweight monitoring agent (141) to ZKP verification (143) and on-chain confirmation. Fig. 7 The profit distribution flowchart of the on-chain settlement smart contract (130) shows the overall settlement flow including the application of a progressive rate (132), a multi-contributor split (133), and a reinvestment conversion (134). Fig. 8 This is a 3-tier license structure diagram that compares the conditions and obligations from the 1st license tier to the 3rd license tier. Fig. 9This is an operation flowchart of the unauthorized integrated detector (160) and the access blocking mechanism (162), showing the sequential flow from violation detection to access blocking. Fig. 10 It is a multi-contributor revenue splitting structure diagram, representing the registration of contributor-specific ratios for co-development Experts and the flow of automatic split remittances. Fig. 11 This is an explanatory diagram of the Softmax Weight-based precision contribution calculation method, and includes a numerical comparison with the simple selection frequency-based method. Fig. 12 This is a flowchart of the operation of the license automatic cancellation module (170), showing the flow from the Expert removal event to the end of the settlement obligation. Specific details for implementing the invention

[0020] 1. Overall System Configuration

[0021] As shown in FIG. 1, the open source Expert license automatic settlement system (100) of the present invention is structured such that a weight public register (110), a license condition automatic detection module (120), an on-chain settlement smart contract (130), a contribution measurement oracle (140), an on-chain registry (150), an unauthorized integration detector (160), a license automatic cancellation module (170), and a decentralized storage (190) are interconnected via a blockchain network (200). An expert module developer terminal (300) registers expert module weights through the weight public register (110), a MoE model operator terminal (400) integrates the corresponding weights into its own MoE model (180) through the license condition automatic detection module (120), and an end user terminal (500) uses the operator's (400) inference service and pays a usage fee. A portion of the usage fee is automatically distributed to the wallet address of the developer terminal (300) by the on-chain settlement smart contract (130).

[0022] 2. Weight public register (110)

[0023] The weight public register (110) serves as an entry point for the expert module developer terminal (300) to publish an expert module weight file that it has independently trained, along with open source license conditions. The weight public register (110) consists of a decentralized storage interface (111), a hash generation module (112), an on-chain registration agent (113), a quality metadata generator (114), and a multi-contributor registration interface (115).

[0024] The decentralized storage interface (111) uploads the expert module weight file received from the developer terminal (300) to a decentralized storage (190), such as IPFS or Arweave, and receives the corresponding storage address (Content Identifier, CID). Since directly storing the entire expert module weight on the blockchain would result in extremely high gas costs due to the nature of large-scale parameter files, the decentralized storage interface (111) adopts an Off-chain Storage, On-chain Proof architecture in which the weight file is stored off-chain in the decentralized storage (190), and only its reference address and cryptographic hash value are recorded on-chain.

[0025] The hash generation module (112) generates a cryptographic hash value by applying the SHA-256 or Keccak-256 algorithm to a weight file uploaded to a decentralized storage (190). This hash value is the unique digital fingerprint of the weight file and has deterministic and collision-resistant characteristics, such that the same hash value is always produced for the same file, and a completely different hash value is generated if the file content is changed by even 1 bit. The cryptographic hash value generated by the hash generation module (112) is recorded in the on-chain registry (150) and subsequently used as a reference value by the license condition automatic detection module (120) to verify the identity of the downloaded weight.

[0026] The on-chain registration agent (113) submits a transaction to the blockchain network (200) containing the cryptographic hash value calculated by the hash generation module (112), the CID returned by the decentralized storage interface (111), the open source license conditions specified by the developer terminal (300), the revenue sharing ratio, the developer's blockchain wallet address, a unique Expert ID, and a registration timestamp, and records it in the on-chain registry (150). The open source license conditions may be GPL-like conditions including a revenue sharing obligation for commercial use, and may consist of multiple tiers (Clause 12), such as allowing free non-commercial use, distributing a certain percentage of inference usage fees for commercial use, or entering into an exclusive usage contract. Since the immutability of the information recorded on-chain is guaranteed by the consensus mechanism of the blockchain network (200), no party can unilaterally change the license conditions or revenue sharing ratio registered subsequently.

[0027] The quality metadata generator (114) records the training dataset source information, the number of training epochs, the validation loss value, and the domain classification tag (e.g., medical, legal, semiconductor design, multilingual translation, etc.) input by the developer terminal (300) together as quality metadata in the on-chain registry (150). The on-chain registry (150) automatically calculates and publishes the quality grade for each expert module based on this quality metadata, and the MoE model operator (400) can efficiently search for expert modules suitable for their MoE model (180) by specifying the domain tag, validation loss value range, quality grade, etc., as composite filter conditions through the search interface.

[0028] The multi-contributor registration interface (115) provides the function of explicitly registering each contributor's blockchain wallet address and contribution ratio in the on-chain registry (150) in the case of an expert module in which multiple researchers and developers, rather than a single developer, have jointly contributed. For example, university research team A (wallet address: 0xABCD...) can be jointly registered with a contribution ratio of 50%, data contributing company B (wallet address: 0xEFGH...) with 30%, and fine-tuning contributor C (wallet address: 0xIJKL...) with 20%. This multi-contributor information is subsequently used as a standard for the multi-contributor split transfer device (133) of the on-chain settlement smart contract (130) to automatically distribute profits.

[0029] 3. On-chain registry (150)

[0030] The on-chain registry (150) is implemented in the form of a smart contract deployed on a blockchain network (200) and includes an Expert ID record (151) generated by a transaction submitted by the weighted public registry (110). The Expert ID record (151) uses the Expert ID, which is a unique identifier of the expert module, as the primary key and includes fields such as a SHA-256 hash value managed by a cryptographic hash storage (152), a CID of a decentralized storage (190), a developer wallet address, license tier information and revenue sharing ratio operated by a license tier manager (153), architecture metadata (embedding dimension, layer structure, normalization method), quality metadata, a list of current integrated operators, and a registration timestamp.

[0031] The License Tier Manager (153) manages multiple license versions for each Expert Module as described in Paragraph 12. The first License Tier stipulates conditions that allow free use for non-commercial research and educational purposes without the obligation of revenue sharing. The second License Tier stipulates conditions that impose an obligation to distribute a pre-agreed ratio of inference usage fees to the original developer when used for commercial inference services. The third License Tier stipulates conditions that, in exchange for granting a specific MoE model operator exclusive rights to use the corresponding Expert Module, simultaneously impose an obligation to distribute inference usage fees along with a lump-sum prepaid license fee upon contract conclusion. The MoE model operator (400) selects one of the three tiers at the time of embedding the weights, and the selection is automatically recorded in the on-chain registry (150) and subsequently used as the basis for settlement in the on-chain settlement smart contract (130).

[0032] The violation registry (154) is a sub-register that accumulates and records violation facts transmitted from the unauthorized integration detector (160) using Expert ID and Operator ID as keys. Since violation records cannot be deleted or modified due to the immutability of the blockchain network (200), the violation history is permanently preserved on-chain.

[0033] 4. License condition automatic detection module (120)

[0034] The license condition automatic detection module (120) is installed in the form of an SDK on the MoE model operator terminal (400) and is automatically activated when the operator downloads expert module weights from a decentralized repository (190) or a public hub (such as Hugging Face Hub). The license condition automatic detection module (120) consists of a hash value comparison engine (121), a license condition parser (122), a commercial use discriminator (123), and an integrated fact recorder (124).

[0035] The hash value comparison engine (121) calculates a hash value by applying the same hash algorithm (SHA-256 or Keccak-256) to the weight file that the operator terminal (400) has completed downloading, and automatically compares this with the value recorded in the cryptographic hash storage (152) of the on-chain registry (150) via the blockchain network (200). If the hash values ​​match, it is cryptographically confirmed that the file is identical to the original registered on-chain, and the license condition parser (122) is activated. If the hash values ​​do not match, it means that the file has been tampered with or is an unregistered weight on-chain, so an event is transmitted to the periodic scanner (161) of the unauthorized integration detector (160).

[0036] When the license condition parser (122) receives a match signal from the hash value comparison engine (121), it parses the license tier information, revenue distribution ratio, developer wallet address, and multiple contributor information of the corresponding Expert from the Expert ID record (151) of the on-chain registry (150) and transmits them to the embed process of the operator terminal (400). The operator checks the parsed license conditions and selects the tier to apply, and this selection is recorded on-chain through the integrated fact recorder (124).

[0037] As described in paragraph 8, the commercial use identifier (123) determines whether the inference service of the operator terminal (400) is operated for commercial purposes by comparing an indicator received from an external oracle (one or more of the number of monthly API calls, number of paid subscribers, and generated revenue) with a pre-set commercial use criterion value. If the determination result is classified as non-commercial use, the first license tier is automatically applied to exempt the revenue sharing obligation, and if it is classified as commercial use, the revenue sharing obligation according to the tier selected by the operator is notified to the on-chain settlement smart contract (130) to activate it.

[0038] The integration fact recorder (124) updates the integration operator list field of the on-chain registry (150) by submitting the operator's weighted integration act, selected license tier, on-chain identifier (Operator ID) of the integrated MoE model, and integration timestamp as a transaction to the blockchain network (200). This record subsequently functions as a legal basis for determining the timing of the occurrence of the profit distribution obligation.

[0039] 5. Expert integration structure of the MoE model (180)

[0040] As illustrated in FIG. 5, the MoE model (180) includes a gating router (181), an internal expert module pool (182), and an external expert module pool (183). The internal expert module pool (182) consists of expert modules that the operator (400) has trained and possesses, and the external expert module pool (183) consists of expert module weights of third-party developers registered on-chain through a weight public register (110) and integrated through an automatic license condition detection module (120). The gating router (181) calculates a softmax score based on the hidden state vector of the input token for each inference request and dynamically selects the top K (Top-K) expert modules from the entire internal expert module pool (182) and external expert module pool (183). Whenever the gating router (181) selects an Expert included in the external expert module pool (183), the lightweight monitoring agent (141) logs the Expert ID, selection softmax weight, inference request identifier, and timestamp to the log.

[0041] If the weights of the external expert module (183) differ from the embedding dimensions of the internal expert module pool (182), the Linear Projection Adapter described in paragraph 4 is automatically generated to perform a linear transformation between the two embedding spaces. The Linear Projection Adapter initially corrects the parameters of the projection matrix using small-scale training data, and this correction process maintains the integrity of the original developer's weights by learning only the adapter layer without modifying the original weights. Since the architecture metadata (embedding dimensions, layer structure, normalization method) is already registered in the on-chain registry (150), the license condition automatic detection module (120) can immediately determine whether automatic adapter generation is necessary at the time of integration.

[0042] 6. Contribution Measurement Oracle (140)

[0043] The contribution measurement oracle (140) consists of a lightweight monitoring agent (141), a commit-rebuild aggregator (142), a ZKP generator (143), and a softmax weight analyzer (144), and is responsible for submitting to the on-chain in a cryptographically verifiable form the frequency and weight of each Expert selected by the gating router (181) of the MoE model (180) from the external expert module pool (183). The core design goal of the contribution measurement oracle (140) is to cryptographically block the possibility of gating log manipulation by the operator side, while simultaneously not exposing sensitive inference data inside the operator server to the outside.

[0044] The lightweight monitoring agent (141) is a software module installed in the form of a sidecar within the inference server process of the operator terminal (400). The lightweight monitoring agent (141) hooks the Expert selection event of the gating router (181) and accumulates the selected Expert ID and the corresponding softmax weight in a memory log buffer in real time for each inference request. The lightweight monitoring agent (141) collects logs asynchronously to minimize the impact on inference performance and aggregates the logs at regular intervals (e.g., 1 hour or 1 day) and transmits them to the commit-rebuild aggregator (142).

[0045] The commit-reveal aggregator (142) processes the aggregation results collected from the lightweight monitoring agent (141) in a two-stage process. In the first stage (Commit), the cryptographic hash value of the entire aggregation result is first submitted to the blockchain network (200) as a transaction to be confirmed on-chain. This commit hash value serves as evidence proving that the actual aggregated data to be disclosed in the subsequent Reveal stage has not been tampered with. After a pre-agreed waiting time (e.g., 24 hours) has elapsed, the actual aggregated data is submitted to the blockchain network (200) in the second stage (Reveal). The on-chain smart contract then automatically compares the commit hash value with the hash value of the revealed data to verify whether they match. If an operator wishes to manipulate the aggregated data after the commit, they would have to change the commit hash value that has already been confirmed on-chain; therefore, under the guarantee of immutability of the blockchain, this is cryptographically impossible.

[0046] The ZKP generator (143) generates a Zero-Knowledge Proof for the aggregate data submitted by the commit-rebuild aggregator (142). Specifically, the ZKP generator (143) generates a ZK-SNARK proof using the proposition that "the aggregate value of Expert contribution calculated from a specific gating log matches the publicly disclosed aggregate result" as a public input and the original log of the actual individual inference request as a private witness. Since the on-chain smart contract verifying this proof can mathematically confirm the accuracy of the aggregate result without accessing the individual inference data, the integrity of the contribution data is cryptographically guaranteed while the operator's inference input / output data is not exposed externally.

[0047] The softmax weight analyzer (144) is responsible for the precise contribution calculation function described in Clause 11. When a gating router (181) simultaneously selects multiple Experts in a Top-K manner in a single inference request, the proportion of each Expert's output contributing to the final response is determined by the softmax score rather than the number of simple selections. The softmax weight analyzer (144) receives the softmax weights for each Expert from the lightweight monitoring agent (141) and calculates an integrated contribution score by multiplying them by the selection frequency. For example, if Expert A is selected 700 times in 1,000 inference requests and its average softmax weight is 0.7, the integrated contribution score for Expert A becomes 490; if Expert B is selected 600 times and its average softmax weight is 0.4, the integrated contribution score becomes 240. This integrated contribution score is provided as the profit distribution weight of the on-chain settlement smart contract (130).

[0048] 7. On-chain settlement smart contract (130)

[0049] The on-chain settlement smart contract (130) is a self-executing smart contract deployed on a blockchain network (200) and consists of a contribution receiving interface (131), a progressive distribution ratio calculator (132), a multi-contributor split transfer device (133), and a profit reinvestment converter (134). The on-chain settlement smart contract (130) is automatically triggered when the MoE model operator (400) transfers the inference service fee to the smart contract address, and sequentially executes the following settlement logic.

[0050] The contribution receiving interface (131) receives the contribution aggregate value per Expert, which is verified by the contribution measurement oracle (140) through the ZKP generator (143) and submitted on-chain. The contribution receiving interface (131) verifies the ZK-SNARK proof included in the received contribution data through the on-chain verifier smart contract and allows the aggregate value as input for settlement calculation only if the proof is valid. If the ZK-SNARK proof verification fails, the execution of the corresponding settlement cycle is suspended and the abnormal situation is recorded as an event.

[0051] The progressive distribution ratio calculator (132) applies different profit distribution ratios in stages according to the monthly cumulative inference contribution count of each Expert, as described in paragraph 6. For example, if the monthly contribution count is less than 10,000, a first distribution ratio (e.g., 5% of the inference fee) is applied; if it is 10,000 or more but less than 100,000, a second distribution ratio (e.g., 7%) is applied; and if it is 100,000 or more, a third distribution ratio (e.g., 10%) is applied. This progressive structure functions as an incentive mechanism that induces developers to continuously improve Experts that meet actual demand by providing higher economic rewards to developers who have trained high-quality Experts that are actually utilized in many inferences.

[0052] The multi-contributor split transfer device (133) refers to the multi-contributor information recorded in the Expert ID record (151) of the on-chain registry (150) and automatically splits the settlement amount attributed to the Expert according to each contributor's contribution ratio. The split amount is immediately transferred atomically to each contributor's blockchain wallet address. Atomic transfer means that the transfer to all contributors succeeds or fails entirely within a single transaction, thereby preventing a situation where the amount is transferred to only some contributors and not paid to others.

[0053] The profit reinvestment converter (134) implements the profit reinvestment automation function described in paragraph 9. If the expert module developer (300) registers in advance in the on-chain registry (150) the ratio to be directly transferred to their wallet address (direct receipt ratio) and the ratio to be converted into computing credits for new expert module training (reinvestment ratio), the profit reinvestment converter (134) automatically swaps the amount corresponding to the reinvestment ratio into credit tokens of a linked decentralized computing marketplace (e.g., Akash Network, Render Network, etc.) at the time of settlement and deposits it into the developer's computing credit account. This mechanism forms an automated R&D cycle that allows the Expert developer to immediately reinvest profits into training the next generation of Experts.

[0054] 8. Unauthorized integrated detector (160)

[0055] The unauthorized integration detector (160) is composed of a periodic scanner (161) and an access blocking mechanism (162) and is responsible for automatically detecting and sanctioning the unauthorized integration of on-chain unregistered weights into the MoE model (180). The periodic scanner (161) calculates the cryptographic hash values ​​of all weight files loaded into the external expert module pool (183) of the current MoE model (180) at regular intervals (e.g., 24 hours) from an operator terminal (400) equipped with the SDK of the license condition automatic detection module (120), and compares them collectively with the values ​​recorded in the cryptographic hash storage (152) of the on-chain registry (150) via the blockchain network (200). If, as a result of comparison, a hash value that is not registered on the chain is found, or if it is registered but the fact of integration through the license condition automatic detection module (120) is not recorded, the periodic scanner (161) classifies the fact as a violation event and submits a transaction to the blockchain network (200) to automatically record the relevant information (time of violation, estimated violation Expert hash value, Operator ID) in the violation registry (154) of the on-chain registry (150).

[0056] The access blocking mechanism (162) executes smart contract logic that automatically rejects on-chain registry (150) access requests originating from the wallet address of a specific Operator ID when the cumulative number of violations of that Operator ID recorded in the violation registry (154) exceeds a preset threshold (e.g., 3 times). This blocks the Operator from downloading other expert module weights registered on-chain or entering into new license agreements at the blockchain level. Since the access blocking status is not lifted without an on-chain governance vote or the explicit consent of the original developer to lift it, it functions as a strong and permanent reputation-based sanction against unauthorized use.

[0057] 9. License automatic cancellation module (170)

[0058] The automatic license cancellation module (170) is automatically activated when a MoE model operator (400) removes a specific external expert module weight from their external expert module pool (183). The SDK of the automatic license condition detection module (120) detects the weight unload event in the external expert module pool (183), and submits the corresponding Expert ID, Operator ID, and removal timestamp as a transaction to the blockchain network (200) to remove the operator from the integrated operator list of the on-chain registry (150). As soon as the on-chain settlement smart contract (130) receives this removal event, it automatically terminates the operator's profit distribution obligation to the corresponding Expert from the time of the removal timestamp. Any unsettled balance incurred prior to the removal time is settled based on the last aggregated data verified by the contribution measurement oracle (140) through the final settlement process of the automatic license cancellation module (170), and then automatically transferred to the original developer's wallet address.

[0059] 10. Example 1 - Registration and Revenue Settlement of Medical Specialized Expert Modules

[0060] Medical AI Research Institute A trained a diagnostic assistance specialized expert module weight (hereinafter "Medical Expert") based on anonymized medical record data accumulated over 5 years. The hidden_dim of the Medical Expert is 4,096, the number of parameters is approximately 7B, and a validation loss value of 0.12 was achieved. A researcher at Research Institute A executes the weight public register (110) through a developer terminal (300).

[0061] The decentralized storage interface (111) uploads the medical expert weight file (about 14GB) to IPFS and returns the CID (QmXyz...). The hash generation module (112) applies SHA-256 to the weight file to produce a hash value (0xABCD1234...). The quality metadata generator (114) records the training dataset source (3 million anonymized medical records), training epoch (50), validation loss value (0.12), and domain tag (medical / diagnostic aid). The on-chain registration agent (113) submits a transaction to the blockchain network (200) containing the Expert ID (expert-med-001), hash value, CID, second license tier (8% distribution of inference usage fees), laboratory A wallet address (0xAAAA...), and co-contributor information (data donation hospital consortium B, 0xBBBB..., contribution ratio 30%) to register it in the on-chain registry (150).

[0062] After six months, medical AI service company C discovers medical experts to enhance medical domain performance while operating an LLaMA-based MoE model (180). When a medical expert file is downloaded from an operator terminal (400) with the license condition automatic detection module (120) activated, a hash value comparison engine (121) calculates the hash value of the downloaded file and automatically compares it with the value in the on-chain registry (150) to confirm a match. The license condition parser (122) parses the second license tier and 8% distribution conditions and presents them to operator C. When operator C confirms agreement, the integration fact recorder (124) records the integration fact on-chain, and the medical expert is added to the external expert module pool (183).

[0063] For one month after the inference service of Enterprise C is launched, the lightweight monitoring agent (141) collects the medical expert selection logs of the gating router (181). After one month of aggregation, the number of medical expert selections is recorded as 450,000, and the average softmax weight is recorded as 0.65. The softmax weight analyzer (144) calculates the integrated contribution score as 450,000 × 0.65 = 292,500. The commit-rebuild aggregator (142) first commits the aggregation result hash value and rebuilds the actual data after 24 hours, and the ZKP generator (143) generates a ZK-SNARK proof and submits it on-chain.

[0064] When the on-chain settlement smart contract (130) receives aggregate data in which the contribution receiving interface (131) has successfully verified the ZK-SNARK proof, the progressive distribution ratio calculator (132) classifies the Medical Expert's monthly contribution count of 450,000 as 100,000 or more and applies a third distribution ratio of 10%. If the monthly inference service fee of Company C is 100 million won, 3.5 million won, which is 10% of 35 million won corresponding to the Medical Expert's share (assumed: 35%) in the sum of all Experts' contribution scores, is determined as the settlement amount attributable to the Medical Expert. The multi-contributor split transfer machine (133) executes automatic atomic transfers to the wallet addresses of Research Institute A (70%, 2.45 million won) and Hospital Consortium B (30%, 1.05 million won), respectively. Since Research Institute A has set the reinvestment ratio to 20% upon registration, the profit reinvestment conversion period (134) automatically converts 490,000 won out of 2,450,000 won into computing credits and accumulates them as the next Expert training resource for Research Institute A.

[0065] 11. Example 2 - Unauthorized Integrated Detection and Access Blocking

[0066] Startup D discovered that the semiconductor design specialized expert module (expert-semi-007) was registered in the on-chain registry (150), bypassed the license condition automatic detection module (120), directly downloaded the corresponding weight, and integrated it into the external expert module pool (183) of its MoE model (180). Although the SDK is installed on Startup D's operator terminal (400), the license agreement step was forcibly skipped, so the on-chain record by the integration fact recorder (124) is missing.

[0067] The periodic scanner (161) calculates the hash value of all weights loaded into Startup D's external expert module pool (183) at a 24-hour interval and compares it with the on-chain registry (150). As a result of the comparison, a discrepancy is detected where the hash value (0xXYZ...) of expert-semi-007 is registered on-chain, but the fact of integration with Startup D's Operator ID (0xDDDD...) is not recorded. The periodic scanner (161) records this violation event in the violation registry (154) and sends an on-chain notification to the wallet address of the original Semiconductor Expert developer (0xEEEE...). If Startup D ignores the notification and two additional unauthorized Expert integrations are detected over 30 days, causing the cumulative number of violations to exceed 3, the access blocking mechanism (162) registers Startup D's wallet address (0xDDDD...) on the blacklist of the on-chain registry (150) and automatically rejects all subsequent weight download requests from that address.

[0068] 12. Example 3 - Expert Integration Between Heterogeneous Architectures

[0069] A small research team F registered a legal-specialized expert module (expert-legal-012) trained on a Mistral-7B base model (hidden_dim: 4,096) in an on-chain registry (150). The architecture metadata records the embedding dimension 4,096, the SwiGLU activation function, and the RMSNorm normalization method. Company G operates an LLaMA-3 70B-based MoE model (180) with a hidden_dim of 8,192 and attempts to integrate expert-legal-012 to strengthen the legal domain.

[0070] The license condition automatic detection module (120) completes hash value comparison upon downloading expert-legal-012, and the license condition parser (122) reads the architecture metadata of the on-chain registry (150) to detect a mismatch between the embedding dimension 4,096 and the embedding dimension 8,192 of the operator MoE model. The license condition automatic detection module (120) determines that automatic generation of a Linear Projection Adapter is required and automatically generates an adapter layer containing a linear projection matrix (W ∈ R^{8192×4096}) of dimensions 4,096→8,192. This adapter layer initially optimizes the parameters of the projection matrix through lightweight calibration using small-scale legal domain data (about 10,000 records), and during the calibration process, the original expert-legal-012 weights are frozen and not modified. Consequently, the gating router (181) of the MoE model (180) is able to integrate the output of the legal expert converted through the adapter layer in the same hidden state space as the internal expert modules.

[0071] 13. Example 4 - Multi-contributor Registration and Commit-Rebuild Settlement

[0072] National AI Research Consortium H registers a multilingual translation expert module (expert-trans-030) jointly trained by three university research teams (I, J, K). Through the multi-contributor registration interface (115), research team I (contribution ratio 40%, wallet address 0xIIII...), research team J (contribution ratio 35%, wallet address 0xJJJJ...), and research team K (contribution ratio 25%, wallet address 0xKKKK...) are jointly registered in the on-chain registry (150). The registered revenue sharing ratio is 6% of the inference usage fee based on the second license tier.

[0073] As a result of global translation service platform L integrating expert-trans-030 into its MoE model (180) and operating it for 3 months, the lightweight monitoring agent (141) records an average monthly selection count of 280,000 for expert-trans-030 and an average softmax weight of 0.72. The commit-rebuild aggregator (142) commits the hash value of the 3-month cumulative aggregation result at the end of the month and rebuilds it the next day, and when the ZK-SNARK proof generated by the ZKP generator (143) passes on-chain verification, the on-chain settlement smart contract (130) is executed. If the total inference usage fee of platform L for 3 months is 300 million won and the contribution ratio of expert-trans-030 is 25% of the total Expert, the attributable settlement amount is 4.5 million won, which is 6% of 75 million won, which is 25% of 300 million won. The progressive distribution ratio calculator (132) classifies the monthly average number of contributions of 280,000 as the third distribution interval and applies a method of adding to the distribution ratio of 6%, and the multi-contributor split transfer machine (133) executes atomic split transfers to the respective wallet addresses of research team I (1.8 million won), research team J (1.575 million won), and research team K (1.125 million won).

Claims

Claim 1 In an automatic revenue settlement system for open-source license-based expert module weights, the system comprises: a weight public register in which an expert module developer registers their expert module weights in a public repository along with open-source license conditions including revenue sharing obligations upon commercial use, and records the cryptographic hash of said weights and the developer's wallet address in an on-chain registry; an automatic license condition detection module that, at the time when a Mixture of Experts (MoE) model operator downloads expert module weights from said public repository and integrates them into the expert pool of their MoE model, queries said on-chain registry to automatically verify the license conditions and revenue sharing ratio of said weights and records the fact of integration; an on-chain settlement smart contract that automatically transfers an amount corresponding to the distribution ratio specified in said license conditions from the usage fees generated in the inference service of said MoE model to the original developer's wallet address in proportion to the actual inference contribution of said expert module; and a contribution measurement oracle that records the selection frequency of gating routers per inference request for each of the plurality of expert modules integrated into said MoE model and utilizes this as a weight for revenue sharing. An open source Expert license automatic settlement system characterized by including Claim 2 An open source Expert license automatic settlement system according to claim 1, characterized in that the license condition automatic detection module further includes an unauthorized integration detection function that detects weights integrated in a state of unregistered license conditions by periodically scanning the cryptographic hash value of the expert module weights mounted in the MoE model and comparing it with the hash value registered in the on-chain registry, and records the fact on-chain.

3. An open source Expert license automatic settlement system according to claim 1, wherein the contribution measurement oracle collects Expert selection logs of a gating router from a lightweight monitoring agent installed on the server of an inference service operator, verifies the integrity of said logs using a Zero-Knowledge Proof, and submits them on-chain to generate reliable contribution data without operator manipulation.

4. An open source Expert license automatic settlement system according to claim 1, wherein the weight disclosure register registers architecture metadata describing the embedding dimension, layer structure, and normalization method of the base model in which the module was trained, along with the expert module weights; and the license condition automatic detection module determines whether to automatically generate a Linear Projection Adapter for expert modules that differ from the operator's MoE model in embedding dimension, by referring to the architecture metadata.

5. An open source Expert license automatic settlement system according to claim 1, wherein the weight disclosure registry records quality metadata on-chain together with the source of the dataset used for training, the number of training epochs, the validation loss value, and a domain classification tag at the time of registration of the expert module weight; and wherein the on-chain registry calculates and discloses a quality grade for each expert module based on the quality metadata, and provides a search interface that allows an MoE model operator to filter expert modules based on the quality grade.

6. An open source Expert license automatic settlement system according to claim 1, wherein the on-chain settlement smart contract sets the profit distribution ratio as a hierarchical structure rather than a single fixed value, and includes a progressive profit distribution structure in which a first distribution ratio is applied when the monthly cumulative inference contribution of the expert module is less than a first threshold, a second distribution ratio higher than the first distribution ratio is applied when the cumulative contribution is greater than or equal to the first threshold but less than a second threshold, and a third distribution ratio higher than the second distribution ratio is applied when the cumulative contribution is greater than or equal to the second threshold.

7. An open source Expert license automatic settlement system, wherein, in addition to an expert module developer registering a single weight, the weight disclosure register further includes a multi-contributor registration function that registers each contributor's wallet address and contribution ratio on-chain for an expert module jointly contributed by multiple developers; and the on-chain settlement smart contract refers to the multi-contributor registration information and automatically divides and remits the revenue attributed to the expert module according to each contributor's contribution ratio.

8. An open source Expert license automatic settlement system according to claim 1, wherein the license condition automatic detection module receives one or more indicators from an external oracle, such as the number of monthly API calls, the number of paid subscribers, or generated revenue of the inference service, in order to determine whether the inference service of the MoE model operator is operated for commercial purposes; and further includes a commercial use determination module that activates the on-chain settlement smart contract to trigger the revenue distribution obligation only when the indicator exceeds a preset commercial use threshold value, and exempts the revenue distribution obligation for non-commercial use below the threshold value.

9. An open source Expert license automatic settlement system according to claim 1, characterized in that the on-chain settlement smart contract further includes a profit reinvestment automation function that, in addition to directly transferring profits to the wallet address of the original developer of the expert module, automatically converts an amount corresponding to a reinvestment ratio pre-specified by the original developer into on-chain computing credits for new expert module training.

10. An open source Expert license automatic settlement system characterized in that, in the second paragraph, the unauthorized integration detection function automatically records the on-chain identifier (Operator ID) of the corresponding MoE model operator in a violation registry when it detects a weight integrated in a state where the license conditions are not registered; and for an operator whose number of violations accumulated in the violation registry exceeds a preset threshold, the system further includes an access blocking mechanism that automatically rejects requests for downloading and integrating other expert module weights registered in the on-chain registry.

11. An open source Expert license automatic settlement system characterized in that, in addition to aggregating the Expert selection logs of the gating router by inference request unit, the contribution measurement oracle utilizes the Softmax Weight, which represents the contribution of each expert module's output value to the final response when multiple expert modules are simultaneously selected for the same inference request, as an additional factor in calculating the contribution, thereby calculating a contribution more precise than simple selection frequency.

12. An open source Expert license automatic settlement system according to claim 1, wherein the on-chain registry supports multiple license versions per expert module, comprising: a first license tier that allows non-commercial use without revenue sharing obligations; a second license tier that imposes an obligation to distribute a certain percentage of inference usage fees for commercial use; and a third license tier that simultaneously imposes a one-time upfront license fee and an obligation to distribute inference usage fees in exchange for granting exclusive usage rights; wherein the MoE model operator selects one of the license tiers at the time of weight embedding, and the selection is automatically recorded in the on-chain registry.

13. An open source Expert license automatic settlement system, wherein, in claim 1, the weight disclosure register adopts an off-chain storage / on-chain proof structure in which, instead of storing the entire expert module weight on-chain, the weight is stored in a decentralized storage (IPFS or Arweave) and only the storage address and cryptographic hash value are recorded in the on-chain registry; and the license condition automatic detection module verifies the integrity and identity of the downloaded weight by comparing the cryptographic hash value of the weight file downloaded by the MoE model operator with the recorded value in the on-chain registry.

14. An open source Expert license automatic settlement system characterized in that, in claim 1, the contribution measurement oracle aggregates gating logs collected from a monitoring agent installed on the server of an inference service operator at regular intervals and submits them on-chain in the form of a batch, and applies a commit-reveal method in which the cryptographic hash value of the aggregate result for each batch is committed on-chain first and the actual aggregated data is disclosed after a certain period of time has elapsed.

15. An open source Expert license automatic settlement system according to claim 1, characterized in that the system further includes an automatic license termination module that, when a MoE model operator removes a specific expert module weight from its Expert pool, automatically records the removal time in an on-chain registry and automatically terminates the revenue distribution obligation of the expert module for inference usage fees occurring thereafter.

Images