Agentic code attestation
AI agents evaluate code for compliance and security in data sharing systems, ensuring trust and collaboration without IP exposure, addressing the limitations of existing cleanroom technologies.
Patent Information
- Application Number
- PCT/US2026/027097
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-07-16
- Filing Date
- 2026-05-07
- Publication Date
- 2026-11-19
AI Technical Summary
Existing data sharing mechanisms, such as cleanrooms, lack the ability to evaluate the safety and compliance of arbitrary code without risking the loss of developers' intellectual property (IP) and require human intervention, which is often not feasible due to technical sophistication limitations.
A system and method using AI agents to evaluate code for adherence to data owner policies without accessing the code, ensuring compliance and security through specialized models that provide attestation without IP exposure.
Enables secure and trustworthy data collaboration by evaluating code for safety and compliance without human intervention, allowing data owners to partner with developers who previously could not due to perceived risks, while protecting IP.
Smart Images

Figure US2026027097_19112026_PF_FP_ABST
Abstract
Description
Attorney Docket No. RAMP-00329-WOAGENTIC CODE ATTESTATION CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] This application claims priority to US Provisional Patent Application No. 63 / 806,101, filed May 15, 2025, and US Provisional Patent Application No. 63 / 844,847, filed July 16, 2025. Such applications are incorporated by reference as it fully set forth herein.BACKGROUND OF THE INVENTION
[0002] When a data owner shares access to its data, it needs to be sure that the data cannot be abused by the data owner's partners who want to run code against the data. Historically, collaborators have relied on contractual mechanisms of trust, such as legal liability assignment, as a retroactive mechanism for providing this assurance. However, in many cases the risk of sharing data was deemed too high to rely on contractual mechanisms alone before executing arbitrary code, and thus profitable opportunities to share data may have been foregone due to the associated risk.
[0003] Data cleanrooms have emerged as a technological mechanism of mitigating the risk of data abuse, which have allowed some of these partnerships to proceed. While cleanrooms provide technical controls ensuring no access to raw data, they are pure infrastructure and do not make a judgment on the actual code executed on shared data within the data cleanroom. In the classic configuration of a cleanroom, a data owner has two choices around allowing code execution or not. The first is to simply trust the code developer. In this case the partner assumes the developer is acting in good faith according to the agreed use cases and contract items and does not need to view the code. This opens the data owner to risk from a malicious code developer. The second choice is to ask to verify the code. In a state of distrust, a data owner will want to ensure that the code does what the developer says it does. There is also a hidden element of trust here, where both parties need to trust the software provider with access to the code dictating what queries are run within a given cleanroom.
[0004] Increasingly, developers are viewing code, especially model weights resulting from training of machine learning (ML) models, as their intellectual property (IP). Given that this IP is their competitive differentiator, they may wish to fiercely protect it against disclosure. This means that they would ideally not allow the data owners or the cleanroom provider to have access to the code. However, data owners want some guarantee that code does what theAttorney Docket No. RAMP-00329-WOdeveloper says it does. The developers do not wish to share code, and even if they did, many data owners wouldn't have the technical sophistication to evaluate it.
[0005] A system and method for allowing data owners or cleanroom providers to evaluate code while not risking loss of the developers' valuable IP, even when the data owner may not have the sufficient technical sophistication to evaluate the code itself, is therefore desired.SUMMARY OF THE INVENTION
[0006] The invention is directed to a system and method for prompting and orchestrating a series of artificial intelligence (Al) agents to evaluate the safety of arbitrary code relative to a data owner's own policies around data access and sharing by other parties. The invention replaces the need for humans to audit and understand arbitrary code before executing it on sensitive datasets. Instead, specialized models are deployed to evaluate code for adherence to compliance regimes, privacy controls, and other security constraints as described in policy inputs from data owners.
[0007] The invention in certain embodiments acts as an impartial, expert evaluator of the code without putting the IP of the code owner at risk. The invention can, in certain embodiments, attest with a particular degree of confidence whether (1) the submitted code does what the code developer says it does; (2) the submitted code complies with the data owner's natural language policies around how its data can be leveraged; (3) the submitted code is not attempting any known malicious activity; (4) the submitted code has not changed since the last evaluation; and / or (5) the evaluation was performed in a transparent manner, with the opportunity for data collaborators to deploy an auditor model. Code integrity between evaluations may be verified through cryptographic hashing of the submitted code, wherein the hash of the submitted code at the time of initial evaluation is stored and compared against a hash of subsequently submitted code.
[0008] In certain embodiments, the invention achieves all of these functions without needing to access source data and without "internalizing" the actual submitted code, such that data owners can make a value judgment on whether to allow code to execute on their data without needing to directly audit the code. The code developers, meanwhile, have assurance that only a machine is evaluating their code, with no fine-tuning, such that there is no risk of IP leakage due to the evaluation being performed. As used herein, "internalizing" means incorporating the submitted code into the model's parameters through training or fine-tuning, or otherwiseAttorney Docket No. RAMP-00329-WOretaining the submitted code in a manner that would allow subsequent reconstruction or extraction of the code from the model.
[0009] In certain embodiments, the invention unlocks the ability for data owners to collaborate with code developer partners with which they previously would not have been allowed to partner given the perceived risk of the collaboration. This form of code attestation brings the ability to audit code even to parties unable to technically evaluate such code, and thus brings trust to situations where such trust could not be achieved due to the technical limitations of the parties.
[0010] These and other features, objects and advantages of the present invention will become better understood from a consideration of the following detailed description of the preferred embodiments and appended claims in conjunction with the drawings as described following: BRIEF DESCRIPTION OF DRAWINGS
[0011] Fig. 1 is an architectural overview of a system according to an embodiment of the present invention.
[0012] Fig. 2 is a diagram of a computing hardware component of an overall distributed architecture for implementation of an embodiment of the present invention.DETAILED DESCRIPTION OF THE INVENTION
[0013] Before the present invention is described in further detail, it should be understood that the invention is not limited to the particular embodiments described, and that the terms used in describing the particular embodiments are for the purpose of describing those particular embodiments only, and are not intended to be limiting, since the scope of the present invention will be limited only by the claims.
[0014] An architectural diagram of an embodiment of the invention is provided in Fig. 1. The embodiment makes use of a number of components that are connected as illustrated.
[0015] The application layer 100 includes a front-end user interface ( U I) with which users interact and through which attestation information and logging is made available to the users. Users include data owner 10 and code developer 12. Data owner 10 may interact with the system through policy setting Ul 102, clean review Ul 106, and attestation log Ul 108. Code developer 12 may interact with the system through code submission Ul 104.
[0016] The orchestration layer 200 provides the technical infrastructure through which the deployment and running of the code attestation framework is managed and invoked.Attorney Docket No. RAMP-00329-WO
[0017] The query execution layer 300 provides the technical infrastructure through which code is executed on data. In some embodiments, the code attestation system operates in conjunction with a secure, multi-party data collaboration cleanroom.
[0018] The auditor runner 202 is a mechanism by which a data owner 10 could select or deploy its own model to observe the code attestation agent's reasoning and report back on trustworthiness.
[0019] The orchestration agent 204 in orchestration layer 200 is a code attestation agent working off of a framework instructing the orchestration layer 200 to invoke the appropriate alternative agents based on the policies set in the policy setting Ul for evaluation. These include policy engine agent 208, code analysis agent 210, privacy agent 212, and intent analysis agent 214.
[0020] The policy engine agent 208, code analysis agent 210, privacy agent 212, and intent analysis agent 214 are specialized agents trained and deployed with models fit for policy evaluation of their specific areas. These agents are task-specific models deployed by the orchestration layer 200 as necessary to evaluate the submitted code against data owner- supplied policies. Other particular agents could be employed in alternative embodiments of the invention.
[0021] The query runner 222 performs the function of, once code is accepted, initiating execution of the code against assigned data owner data 16.
[0022] The attestation runner 220 is a mechanism via which the provider of the system is able to monitor adherence to policies submitted in the policy setting Ul 102. External partners leveraging the results of query execution must interact with the attestation runner 220 on an ongoing basis when accessing the query results 304 in order to attest their use is adhering to demonstrated policies. Data owners 10 and code developers 12 can monitor usage traffic via the attestation log Ul 108.
[0023] Continuing with reference to Fig. 1, a more specific workflow for the system illustrated therein may be now described. The data owner 10 selects default policies or applies natural language policies to apply to a given dataset of data owner data 16 for evaluation via the policy setting Ul 102 in application layer 100. The data owner 10 may optionally select an auditor model through auditor runner 202 to act as an observer to the reasoning of orchestrator agent 204 in performing the code evaluation of developer code 14.Attorney Docket No. RAMP-00329-WO
[0024] The code developer 12 then indicates that it would like to run the developer code 14 on the data owner data 16 via the code submission III 104. The orchestrator agent 204 takes the policy settings and developer code 14 as inputs and routes these assets to the appropriate taskspecific models for evaluation. These include policy engine agent 208, code analysis agent 210, privacy agent 212, and intent analysis agent 214. Each agent has access to a set of software tools (tools repository) 206 that may be used to better perform evaluations. Tools are deployed locally, and no external calls are made, in certain embodiments.
[0025] The policy engine agent 208 parses and classifies the input policies. Depending on the classifications of policies, the orchestrator agent 204 then routes the developer code 14 for review by each deployed task-specific agent. Each agent performs an independent evaluation of the code based on its expertise informed by the tasks it was trained to perform. This evaluation can occur in the context of a virtual private cloud (VPC) 302 hosted by the cleanroom provider, or in the context of a trusted execution environment (TEE) with all of the protections confidential computing provides. The difference between these two modes is that one allows a cleanroom provider machine to potentially access the developer code 14, while the other provides additional guarantees that the cleanroom provider cannot access the developer code 14 under any circumstances.
[0026] Each agent then reports back to the orchestrator agent 204 at report generation agent 216, which synthesizes the evaluation results into an attestation report 218 with a confidence score of adherence of the code to the policies set by the data owner 10. The data owner 10 reviews the attestation report 218 and determines whether or not to proceed with code execution through clean review III 106.
[0027] If the data owner 10 proceeds, the cleanroom 302 is spun up to perform query execution in-memory. Code and data are temporarily brought into the cleanroom virtual private cloud (VPC) 302 or orchestrated TEE on selected hardware for execution through the operation of query runner 222. Query results 304 are written to the cleanroom data plane (query execution layer) 300 and a call is made to the attestation runner 220 to confirm completion in accordance with the policies.
[0028] If results are to be delivered to an external partner, a call is then made to the attestation runner 220 to verify the delivery of results. Upon usage in the partner activation endpoint 306, the partner makes a call to the attestation runner 220 to validate usage in line with data owner 10 and code developer 12 policies.Attorney Docket No. RAMP-00329-WO
[0029] Logs of attestation handshakes between partners collaborating in cleanrooms and activation endpoint partners are presented to users on a rolling basis. Users are notified of any anomalous usage or changes in code or attestation calls, i.e., in the attestation handshakes that are logged.
[0030] Fig. 2 is a block diagram illustrating an example computer hardware system, according to various embodiments, that can be used as part of a multi-component system to implement the invention as has been described in a cloud-computing environment. Computer system 440 may implement a hardware portion of a cloud computing system as forming parts of the various implementations of the present invention. Computer system 440 may be any of various types of hardware devices, including, but not limited to, a commodity server, personal computer system, desktop computer, laptop or notebook computer, mainframe computer system, handheld computer, workstation, network computer, a consumer device, application server, physical storage device, telephone, mobile telephone, or in general any type of computing node, compute node, compute device, and / or hardware computing device.
[0031] Computer system 440 includes one or more hardware processors 440a, 441b...441n (any of which may include multiple processing cores, which may be single or multi-threaded) coupled to a physical system memory 442 via an input / output (I / O) interface 444. Computer system 440 further may include a network interface 446 coupled to I / O interface 444. In various embodiments, computer system 440 may be a single processor system including one hardware processor 441a, or a multiprocessor system including multiple hardware processors 441a, 441b...441n. Processors 441a, etc. may be any suitable processors capable of executing computing instructions. For example, in various embodiments, processors 441a, etc. may be general-purpose or embedded processors implementing any of a variety of instruction set architectures.
[0032] In multiprocessor systems, each of processors 441a, etc. may commonly, but not necessarily, implement the same instruction set. The computer system 440 also includes one or more hardware network communication devices (e.g., network interface 446) for communicating with other systems and / or components over a communications network, such as a local area network, wide area network, or the Internet. For example, a client application executing on system 440 may use network interface 446 to communicate with a server application executing on a single hardware server or on a cluster of hardware servers that implement one or more of the components of the systems described herein in a cloudAttorney Docket No. RAMP-00329-WOcomputing environment as implemented in various sub-systems. In another example, an instance of a server application executing on computer system 440 may use network interface 446 to communicate with other instances of an application that may be implemented on other computer systems.
[0033] In the illustrated embodiment, computer system 440 also includes one or more physical persistent storage devices 448 and / or one or more I / O devices 450. In various embodiments, persistent storage devices 448 may correspond to disk drives, tape drives, solid-state memory or drives, other mass storage devices, or any other persistent storage devices. Computer system 440 (or a distributed application or operating system operating thereon) may store instructions and / or data in persistent storage devices 448, as desired, and may retrieve the stored instructions and / or data as needed. For example, in some embodiments, computer system 440 may implement one or more nodes of a control plane or control system, and persistent storage 448 may include the solid-state drives (SSDs) attached to that server node. Multiple computer systems 440 may share the same persistent storage devices 448 or may share a pool of persistent storage devices, with the devices in the pool representing the same or different storage technologies, including such technologies as described above.
[0034] Computer system 440 includes one or more physical system memories 442 that may store code / instructions 443 and data 445 accessible by processor(s) 441a, etc. The system memories 442 may include multiple levels of memory and memory caches in a system designed to swap information in memories based on access speed, for example. The interleaving and swapping may extend to persistent storage devices 448 in a virtual memory implementation, where memory space is mapped onto the persistent storage devices 448. The technologies used to implement the system memories 442 may include, by way of example, static randomaccess memory (RAM), dynamic RAM, read-only memory (ROM), non-volatile memory, solid- state memory, or flash-type memory. As with persistent storage devices 448, multiple computer systems 440 may share the same system memory systems 442 or may share a pool of system memories 442. System memory or memory systems 442 may contain program instructions 443 that are executable by processor(s) 441a, etc. to implement the routines described herein.
[0035] In various embodiments, program instructions 443 may be encoded in binary, Assembly language, any interpreted language such as Java, compiled languages such as C / C++, or in any combination thereof; the particular languages given here are only examples. In someAttorney Docket No. RAMP-00329-WOembodiments, program instructions 443 may implement multiple separate clients, server nodes, and / or other components.
[0036] In some implementations, program instructions 443 may include instructions executable to implement an operating system (not shown), which may be any of various operating systems, such as UNIX, LINUX, Solaris™, MacOS™, or Microsoft Windows™. Any or all of program instructions 443 may be provided as a computer program product, or software, that may include a non-transitory computer-readable storage medium having stored thereon instructions, which may be used to program a computer system (or other electronic devices) to perform a process according to various implementations.
[0037] A non-transitory computer-readable storage medium may include any mechanism for storing information in a form (e.g., software or processing application) readable by a machine (e.g., a physical computer). Generally speaking, a non-transitory computer-accessible medium may include computer-readable storage media or memory media such as magnetic or optical media, e.g., disk or DVD / CD-ROM, coupled to or in communication with computer system 440 via I / O interface 444. A non-transitory computer-readable storage medium may also include any volatile or non-volatile media such as RAM or ROM that may be included in some embodiments of computer system 440 as system memory 442 or another type of memory. In other implementations, program instructions may be communicated using optical, acoustical or other form of propagated signals (e.g., carrier waves, infrared signals, digital signals, etc.) conveyed via a communication medium such as a network and / or a wired or wireless link, such as may be implemented via network interface 606. Network interface 446 may be used to interface with other devices 442, which may include other computer systems or any type of external electronic device.
[0038] In some embodiments, system memory 442 may include data store 445, as described herein. In general, system memory 442 and persistent storage 448 may be accessible on other devices 442 through a network and may store data blocks, replicas of data blocks, metadata associated with data blocks, and / or their state, database configuration information, and / or any other information usable in implementing the routines described herein.
[0039] In one embodiment, I / O interface 444 may coordinate I / O traffic between processors 441a, etc., system memory 442, and any peripheral devices in the system, including through network interface 446 or other peripheral interfaces. In some embodiments, I / O interface 444 may perform any necessary protocol, timing or other data transformations to convert dataAttorney Docket No. RAMP-00329-WOsignals from one component (e.g., system memory 442) into a format suitable for use by another component (e.g., processors 441a, etc.). In some embodiments, I / O interface 444 may include support for devices attached through various types of peripheral buses, such as a variant of the Peripheral Component Interconnect (PCI) bus standard or the Universal Serial Bus (USB) standard, as examples. Also, in some embodiments, some or all of the functionality of I / O interface 444, such as an interface to system memory 442, may be incorporated directly into processor(s) 441a, etc.
[0040] Network interface 446 may allow data to be exchanged between computer system 440 and other devices attached to a network, such as other computer systems (which may implement one or more storage system server nodes, primary nodes, read-only node nodes, and / or clients of the database systems described herein), for example. In addition, I / O interface 444 may allow communication between computer system 440 and various I / O devices 450 and / or remote storage 448. Input / output devices 450 may, in some embodiments, include one or more display terminals, keyboards, keypads, touchpads, scanning devices, voice or optical recognition devices, or any other devices suitable for entering or retrieving data by one or more computer systems 440. These may connect directly to a particular computer system 440 or generally connect to multiple computer systems 440 in a cloud computing environment, grid computing environment, or other system involving multiple computer systems 440.
[0041] Multiple input / output devices 450 may be present in communication with computer system 440 or may be distributed on various nodes of a distributed system that includes computer system 440. In some embodiments, similar input / output devices may be separate from computer system 440 and may interact with one or more nodes of a distributed system that includes computer system 440 through a wired or wireless connection, such as over network interface 446.
[0042] Network interface 446 may commonly support one or more wireless networking protocols (e.g., Wi-Fi / I EEE 802.11, or another wireless networking standard). Network interface 446 may support communication via any suitable wired or wireless general data networks, such as other types of Ethernet networks, for example. Additionally, network interface 446 may support communication via telecommunications / telephony networks such as analog voice networks or digital fiber communications networks, via storage area networks such as Fibre Channel SANs, or via any other suitable type of network and / or protocol. In variousAttorney Docket No. RAMP-00329-WOembodiments, computer system 440 may include more, fewer, or different components (e.g., displays, video cards, audio cards, peripheral devices, or an Ethernet interface).
[0043] Any of the distributed system embodiments described herein, or any of their components, may be implemented as one or more network-based services in the cloud computing environment. For example, a read-write node and / or read-only nodes within the database tier of a hardware database system may present database services and / or other types of physical data storage services that employ the distributed storage systems described herein to clients as network-based services.
[0044] In some embodiments, a network-based service may be implemented by a software and / or hardware system designed to support interoperable machine-to-machine interaction over a network. A web service may have an interface described in a machine-processable format. Other systems may interact with the network-based service in a manner prescribed by the description of the network-based service's interface. For example, the network-based service may define various operations that other systems may invoke, and may define a particular application programming interface (API) to which other systems may be expected to conform when requesting the various operations.
[0045] In various embodiments, a network-based service may be requested or invoked through the use of a message that includes parameters and / or data associated with the network-based services request. Such a message may be formatted according to a particular markup language such as Extensible Markup Language (XML), and / or may be encapsulated using a protocol. To perform a network-based services request, a network-based services client may assemble a message including the request and convey the message to an addressable endpoint (e.g., a Uniform Resource Locator (URL)) corresponding to the web service, using an Internet-based application layer transfer protocol such as Hypertext Transfer Protocol (HTTP).
[0046] Unless otherwise stated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
[0047] Although any methods and materials similar or equivalent to those described herein can also be used in the practice or testing of the present invention, a limited number of the exemplary methods and materials are described herein. It will be apparent to those skilled in the art that many more modifications are possible without departing from the inventive concepts herein.Attorney Docket No. RAMP-00329-WO
[0048] All terms used herein should be interpreted in the broadest possible manner consistent with the context.
[0049] When a grouping is used herein, all individual members of the group and all combinations and sub-combinations possible of the group are intended to be individually included.
[0050] When a range is stated herein, the range is intended to include all sub-ranges within the range, as well as all individual points within the range.
[0051] When "about," "approximately," or like terms are used herein, they are intended to include amounts, measurements, or the like that do not depart significantly from the expressly stated amount, measurement, or the like, such that the stated purpose of the apparatus or process is not lost.
[0052] All references cited herein are hereby incorporated by reference to the extent that there is no inconsistency with the disclosure of this specification.
[0053] The present invention has been described with reference to certain preferred and alternative embodiments that are intended to be exemplary only and not limiting to the full scope of the present invention, as set forth in the appended claims.
Claims
Attorney Docket No. RAMP-00329-WOClaims1. A code attestation system comprising:a policy interface configured to receive, from a data owner, one or more policies governing permissible execution on data owner data of third-party code;an orchestration agent configured to receive submitted code from a code developer and to route the submitted code and the policies to one or more specialized artificial intelligence (Al) agents for evaluation;a plurality of specialized Al agents, each configured to independently evaluate the submitted code within a respective area of expertise and to return an evaluation result to the orchestration agent;a report generation agent configured to synthesize the evaluation results from the plurality of specialized Al agents into an attestation report that includes a confidence score indicating a degree of adherence of the submitted code to the policies; anda query execution environment configured to execute the submitted code against the data owner data upon approval of the attestation report by the data owner and to produce query results from such execution,wherein the attestation report is generated without internalizing the submitted code and without exposing the submitted code to the data owner.
2. The system of claim 1, wherein the plurality of specialized Al agents comprises at least two of:a policy engine agent configured to parse and classify the policies and to provide classification outputs to the orchestration agent to inform routing of the submitted code to one or more of the specialized Al agents;a code analysis agent configured to evaluate structure and behavior of the submitted code; a privacy agent configured to assess compliance of the submitted code with privacy controls; andan intent analysis agent configured to determine the intent of the submitted code.Attorney Docket No. RAMP-00329-WO3. The system of claim 1, further comprising an auditor runner configured to enable the data owner to select or deploy an auditor Al model that observes reasoning of the orchestration agent during evaluation of the submitted code and reports on trustworthiness of the evaluation.
4. The system of claim 1, wherein each of the plurality of specialized Al agents has access to a locally deployed tools repository, and wherein evaluation operations are performed without external network calls.
5. The system of claim 1, wherein the query execution environment comprises one of:a virtual private cloud (VPC) cleanroom; ora trusted execution environment (TEE) configured to provide confidential computing guarantees such that the submitted code is inaccessible to a cleanroom provider during execution.
6. The system of claim 1, further comprising an attestation runner configured to monitor adherence to the policies during and after execution of the submitted code.
7. The system of claim 1, wherein the attestation report is configured to indicate one or more of:whether the submitted code performs a function described by the code developer; whether the submitted code complies with the policies;whether the submitted code contains or attempts known malicious activity;whether the submitted code has been modified since a prior evaluation; andwhether the evaluation was performed in a transparent manner subject to auditing.
8. A computer-implemented method for attesting code safety without exposing code intellectual property, comprising:receiving, via a policy interface, one or more policies from a data owner specifying constraints on how the data owner's data may be accessed or used by third-party code;receiving submitted code from a code developer seeking to execute the code against the data owner's data;routing, by an orchestration agent, the submitted code and the policies to a plurality of specialized artificial intelligence (Al) agents;Attorney Docket No. RAMP-00329-WOperforming, by each specialized Al agent independently, an evaluation of the submitted code within the agent's respective area of expertise;synthesizing, by a report generation agent, the evaluations into an attestation report that includes a confidence score of adherence of the submitted code to the policies;presenting the attestation report to the data owner for approval, without disclosing the submitted code to the data owner; andupon receiving approval from the data owner, executing the submitted code against the data owner's data in a secure execution environment to produce query results.
9. The method of claim 8, wherein receiving the one or more policies comprises receiving natural language policies entered by the data owner, and wherein the policies are parsed and classified by a policy engine agent of the plurality of specialized Al agents.
10. The method of claim 8, further comprising:receiving, from the data owner, a selection of an auditor Al model; anddeploying the auditor Al model to observe and report on the reasoning of the orchestration agent during evaluation of the submitted code.
11. The method of claim 8, wherein routing the submitted code comprises:classifying, by a policy engine agent, the policies to identify policy classifications corresponding to applicable evaluation areas; andselectively routing the submitted code to one or more of a code analysis agent, a privacy agent, and an intent analysis agent based on the policy classifications identified in the classifying step.
12. The method of claim 8, wherein performing independent evaluations comprises accessing, by each specialized Al agent, a locally deployed tools repository, without making calls to external services or networks.
13. The method of claim 8, wherein executing the submitted code comprises:temporarily loading the submitted code and the data owner's data into the secure execution environment comprising one of a virtual private cloud (VPC) cleanroom or a trusted execution environment (TEE);Attorney Docket No. RAMP-00329-WOperforming query execution in-memory within the secure execution environment; writing query results to a data plane; andcalling an attestation runner to confirm completion in accordance with the policies.
14. The method of claim 8, further comprising:monitoring, via an attestation runner, adherence to the policies on an ongoing basis following execution of the submitted code; andverifying, via the attestation runner, delivery of query results to one or more external partners.
15. The method of claim 8, further comprising:logging attestation handshakes occurring between collaborating parties and partner activation endpoints; andnotifying users of anomalous usage patterns or changes in the submitted code or attestation calls.
16. A non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the processors to perform operations comprising:receiving one or more data owner policies specifying constraints on permissible use of data owner data by third-party code;receiving submitted code from a code developer;orchestrating a plurality of specialized artificial intelligence (Al) agents to independently evaluate the submitted code against the policies, wherein each specialized Al agent operates within a respective area of expertise;generating an attestation report based on evaluation results from the plurality of specialized Al agents, the attestation report including a confidence score indicating a degree of policy adherence of the submitted code; andconditionally executing the submitted code against the data owner data in a secure execution environment based on data owner approval of the attestation report,Attorney Docket No. RAMP-00329-WOwherein the submitted code is not internalized by any model participating in the evaluation and is not disclosed to the data owner.
17. The non-transitory computer-readable medium of claim 16, wherein the operations further comprise deploying an auditor Al model, selected by the data owner, to observe reasoning of an orchestration agent coordinating the plurality of specialized Al agents, and to generate a trustworthiness report on the evaluation.
18. The non-transitory computer-readable medium of claim 16, wherein the secure execution environment comprises a trusted execution environment (TEE) configured to provide confidential computing guarantees such that the submitted code and data owner data are inaccessible to any party outside the TEE during execution.
19. The non-transitory computer-readable medium of claim 16, wherein the operations further comprise:monitoring, via an attestation runner, adherence to the policies throughout execution of the submitted code and delivery of query results to external partners;validating, at a partner activation endpoint, each external usage of query results against applicable policies; andgenerating logs of all attestation handshakes for review by the data owner and the code developer.
