# Table of Contents - [Welcome to Eidon | Eidon AI](#welcome-to-eidon-eidon-ai) - [Post Data | Eidon AI](#post-data-eidon-ai) - [Explore Quests | Eidon AI](#explore-quests-eidon-ai) - [MemAgents | Eidon AI](#memagents-eidon-ai) - [Points Grounding | Eidon AI](#points-grounding-eidon-ai) - [Technical Details | Eidon AI](#technical-details-eidon-ai) - [Developer API | Eidon AI](#developer-api-eidon-ai) - [Transactions & Events | Eidon AI](#transactions-events-eidon-ai) - [Future Tokenomics | Eidon AI](#future-tokenomics-eidon-ai) - [Slashing Reasons | Eidon AI](#slashing-reasons-eidon-ai) - [Request Data | Eidon AI](#request-data-eidon-ai) - [Sample Importance | Eidon AI](#sample-importance-eidon-ai) - [Outlier Likelihood | Eidon AI](#outlier-likelihood-eidon-ai) - [Post Quality MemAgent | Eidon AI](#post-quality-memagent-eidon-ai) - [Eidon State | Eidon AI](#eidon-state-eidon-ai) - [Evaluating VLM MemAgents | Eidon AI](#evaluating-vlm-memagents-eidon-ai) - [User Quality (UQ) MemAgent | Eidon AI](#user-quality-uq-memagent-eidon-ai) - [Reward Accrual | Eidon AI](#reward-accrual-eidon-ai) - [Alignment Score | Eidon AI](#alignment-score-eidon-ai) - [MVP Rationale | Eidon AI](#mvp-rationale-eidon-ai) - [Violation Tolerance | Eidon AI](#violation-tolerance-eidon-ai) - [Post Quality ( PQ ) Score | Eidon AI](#post-quality-pq-score-eidon-ai) - [Slashing Mechanism | Eidon AI](#slashing-mechanism-eidon-ai) - [Reward Mechanism | Eidon AI](#reward-mechanism-eidon-ai) - [Decentralization Path | Eidon AI](#decentralization-path-eidon-ai) - [Dispute & Guardian System | Eidon AI](#dispute-guardian-system-eidon-ai) - [Get Rewarded | Eidon AI](#get-rewarded-eidon-ai) - [Eidon FAQ | Eidon AI](#eidon-faq-eidon-ai) - [Strikes and Bans | Eidon AI](#strikes-and-bans-eidon-ai) - [Robustness to Sybil Attacks | Eidon AI](#robustness-to-sybil-attacks-eidon-ai) - [Email Protection | Cloudflare](#email-protection-cloudflare) --- # Welcome to Eidon | Eidon AI [NextExplore Quests](/get-involved/explore-quests) Last updated 4 months ago Eidon is a decentralized AI network where human-AI agents collaborate to capture real-world data, training the next generation of multimodal AI models. Built on blockchain principles, Eidon ensures contributors own and benefit from their data, creating a permissionless, community-driven AI ecosystem. #### [](#interested-in-going-deeper-lets-dive-in) **Interested in going deeper? Let's dive in:** ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252Fgit-blob-5df7cc6bb0818ed86a508691c2c241b882c55d88%252Fwelcome.png%3Falt%3Dmedia&width=768&dpr=4&quality=100&sign=c98c001&sv=2) --- # Post Data | Eidon AI [PreviousExplore Quests](/get-involved/explore-quests) [NextGet Rewarded](/get-involved/rewards-for-submitting-quests) Last updated 6 months ago You can access Quests and post data through any Eidon Application. The is our main application currently, with more coming soon. #### [](#posting-data-via-the-eidon-app-on-mobile-or-browser) Posting Data via the (on mobile or browser) * Install the app on your home page and log in * Navigate to the Quests page and pick whichever you prefer * Tap on the capture button for your chosen modality at the bottom of the page to capture data. * Alternatively, from anywhere else in the app, you can tap the `+` button to capture any data and select the Quest you want to post to. Keep in mind that different Quests accept different modalities. **Choose your adventure and jump in!** [Eidon App](https://app.eidon.ai/) [Eidon App](https://app.eidon.ai/) [![Logo](https://app.eidon.ai/icons/touch-icon-iphone-retina.png)Eidon App](https://app.eidon.ai/) ![Page cover image](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FGXaJ99HsTDDTg1J2Wtx5%252Feidon_2.webp%3Falt%3Dmedia%26token%3Df78db968-00ac-4bbb-89c4-5a48dd7e6e69&width=1248&dpr=4&quality=100&sign=8e050cba&sv=2) --- # Explore Quests | Eidon AI [PreviousWelcome to Eidon](/) [NextPost Data](/get-involved/post-data) Last updated 4 months ago A Quest is a specific request for data by Eidon or a partner company (see here for more information about creating a Quest for your use case). Users can submit data to Quests to earn rewards. Each Quest has specific instructions for the data they are seeking. Our reward system is tuned to rate incoming data based on these requirements. Users need to follow the Quest instructions and guidelines to ensure the highest possible rewards. Quests also contain examples of high-quality data, the accepted modalities (video, image, audio, etc), **maximum possible** rewards users can earn **per post**, and other helpful information. Think of Quests as guided missions to gather the next-generation of frontier data needed to continue advancing AI. You can see all the available Quests below: [![Logo](https://app.eidon.ai/icons/touch-icon-iphone-retina.png)Eidon App](https://app.eidon.ai/quests) ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FwacAQ50BfkV5wT37s1ES%252FScreenshot%25202024-10-01%2520at%252010.36.25%25E2%2580%25AFAM.jpeg%3Falt%3Dmedia%26token%3Df79fc3ad-de8d-4afa-b925-ce554498e795&width=768&dpr=4&quality=100&sign=46329769&sv=2) ![Page cover image](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FJLcKfJFiZEtoXfayhjJg%252Feidon_1.webp%3Falt%3Dmedia%26token%3Ddb68cc2e-52bc-4300-a23c-2f1966b4d99e&width=1248&dpr=4&quality=100&sign=36337e50&sv=2) --- # MemAgents | Eidon AI [PreviousTransactions & Events](/the-network/transactions-and-events) [NextReward Mechanism](/the-network/reward-mechanism) Last updated 6 months ago MemAgents are programmable actions that handle multiple aspects of the network. MemAgent actions change the , ensuring the integrity of Eidon's canonical record of contributions. Currently, we have the following MemAgents on the network: * **Data Post Agent:** Creates the Metadata objects for each user post and writes it to the sequencer. * **Post Quality Agent:** Runs our algorithms to measure the quality of a given post. * **Reward Agent:** Calculates the rewards for each user's uploads based on the post quality and user quality * **Payment Agent:** Performs balance accounting, paying out accrued rewards every 12 hours. * **User Quality Agent:** Updates the user reputation system every 12 hours based on previous uploads. MemAgents can be triggered by time, user actions, or specific on-chain / off-chain events. We will continue to add more MemAgents over the next months, covering multiple aspects from data usage to other contributions to the network. ### [](#key-memagent-categories) **Key MemAgent Categories:** * **Reward Management:** Distributes rewards to data providers based on data usage and contributions. * **Dataset Curation:** Creates and maintains high-quality datasets by grouping similar data pieces and managing data inclusion. * **Data Quality Assurance:** Assesses the quality of submitted data and user contributions, ensuring the reliability and value of data uploaded. [State of the network](/the-network/eidon-state) [Reward Mechanism](/the-network/reward-mechanism) --- # Points Grounding | Eidon AI [PreviousMVP Rationale](/the-network/reward-mechanism/mvp-rationale) [NextTechnical Details](/the-network/reward-mechanism/technical-details) Last updated 6 months ago It is tricky to determine precisely the value of data. A single video may be worth wildly different amounts based on multiple factors such as length, language, subject, etc. Further, until now, no large scale open market has emerged for data. Data deals happen behind closed doors, in bilateral agreements between corporations, making it impossible for natural market dynamics to set the appropriate price for data. Still, the Eidon team extensively surveyed existing data deals to determine a grounding system for our points. This is not an absolute value for data types, but a **proportional value scale** for different **data modalities**. [](#modality-scale) Modality Scale --------------------------------------- Modality Value Paragraph of Text 1x Image 10x Audio (Speech with Texture) 25x Audio (Task Oriented Dialogue) 50x Audio (Non-Speech Events) 75x Video 100x The table above varies based on `X`, which corresponds to the value of the **data category**. For instance, videos of sporting events are less valuable than POV videos of driving, which are less valuable than videos of task completions. The differential value of categories is harder to quantify, as it requires real-time knowledge of demand and supply in the data market - which is often achievable only through the collective intelligence of open markets. Yet, our system accounts for different `X` values, setting differential category point prices per Quest. This `X` value is implicit in the max value of points per modality displayed in every Quest as seen below. [Quest](/get-involved/explore-quests) ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252Fj0D0M42PYj6NDvx0h7vn%252FScreenshot%25202024-10-02%2520at%252011.59.35%25E2%2580%25AFAM.png%3Falt%3Dmedia%26token%3Dc28283e3-4cff-4781-9adb-fbcadb443c75&width=300&dpr=4&quality=100&sign=f6e55df4&sv=2) ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FNFGM1gblsZUf3uLhtf3M%252FScreenshot%25202024-10-02%2520at%252011.59.56%25E2%2580%25AFAM.png%3Falt%3Dmedia%26token%3D152326c4-2cdc-47d3-be5d-475f658164ca&width=300&dpr=4&quality=100&sign=33f3aef3&sv=2) --- # Technical Details | Eidon AI In this setup, the core components : **instruction** and **data** serve as the primary anchors for **aligning contributions with the goals of the decentralized network**. Instructions represent the specific guidelines or objectives that users must follow to create valuable content. These may be set by the community, AI models, or platform standards, and they define the overarching goals of contributions. By clearly stating what is expected, instructions provide a framework for evaluating user content in alignment with the network's needs. Instructions could range from technical guidelines (e.g., format, quality standards) to thematic directives (e.g., focusing on certain topics) and ethical guidelines (e.g., fairness, relevance). Users respond to instructions by uploading data that they believe satisfies the requirements. This data varies in form (text, images, datasets, etc.) but is assessed uniformly against the instruction’s criteria. At a high level, the reward system is structured around a **collaborative network of specialized AI agents**. Each of these AI agents is initially trained to perform a specific task (often within a particular modality such as image, video, audio, or text), enabling them to effectively process and evaluate data in their area of expertise. By combining the capabilities of these agents, the reward system leverages a diverse skill set across multiple modalities, forming a mechanism for understanding and scoring the user-submitted content. We refer to the resulting scoring agent as Post Quality (PQ) MemAgent. In addition to evaluating contributions, the reward system incorporates a User Quality (UQ) MemAgent that tracks and assesses long-term user behavior to build a reputation score for each user. Unlike individual content assessments, this reputation system enables the reward mechanism to recognize and reward users who consistently demonstrate high standards, originality, and alignment with network goals. The PQ score and UQ score are combined with the task value to compute a final reward score of the uploaded data -- reflecting the contribution of the data point towards DecAI. In the following sections, we break down the components of the point system into key areas that assess content uniqueness, quality, and alignment. [PreviousPoints Grounding](/the-network/reward-mechanism/points-grounding) [NextPost Quality MemAgent](/the-network/reward-mechanism/technical-details/post-quality-memagent) Last updated 5 months ago --- # Developer API | Eidon AI ### [](#building-the-future-of-data) Building the Future of Data While we're actively developing and expanding its capabilities, some features outlined below will be rolled out progressively. **Current Features and Upcoming Developments:** * **Quest Integration (In Progress):** We're actively working on API endpoints to programmatically interact with the Quest system for requesting and fulfilling data requests. Our roadmap includes functionality to set parameters, define pricing, and manage the data acquisition lifecycle. This will empower you to automate data collection and contribute to a thriving data ecosystem. * **Data Access and Model Integration (Early Stages):** Access to datasets and protocol-native models via the API is in its early stages. We envision a standardized interface for building AI-powered applications, refining models, and contributing to model improvements. Stay tuned for updates on this functionality. * **Permissionless Client Expansion (Available):** Our open API and SDK are available now, enabling you to build custom clients and integrations. Expand the Eidon ecosystem and be a part of its growth. This allows for tailored user experiences and a wider range of data applications. * **Decentralized Storage Integration (Future Goal):** While currently centralized, our architecture is designed for a future transition to decentralized storage. The API will adapt as we implement partitioned and fully decentralized storage models, ensuring long-term data integrity and availability for your applications. **Phased Approach to Decentralization:** We're committed to a phased approach to decentralization. The API is designed to evolve alongside the network as we transition to more decentralized models, ensuring a stable and secure experience for developers. **Getting Started:** Explore our current documentation to start building on the Eidon Network. Join our community to stay informed about updates, new features, and contribute to the future of decentralized data. Contact our team with questions or for assistance. [PreviousDispute & Guardian System](/the-network/dispute-and-guardian-system) [NextDecentralization Path](/the-network/path-to-decentralization) Last updated 6 months ago --- # Transactions & Events | Eidon AI [PreviousEidon State](/the-network/eidon-state) [NextMemAgents](/the-network/memagents) Last updated 6 months ago Events and Transactions are the fundamental units of state change within the Eidon network. Each event is executed and recorded on the sequencer - and later on the blockchain - creating a permanent and auditable history of all activity and contributions to the network. ### [](#key-event-categories) **Key Event Categories:** * **Data Management:** Handles the submission, curation, and management of data. * **Example:** Submitting new data, adding data to datasets, deleting data. * **Reward & Payment:** Manages the distribution and claiming of rewards for data providers. * **Example:** Reward accrual transactions, payment transactions. * **Dataset Interactions:** Facilitates the creation, usage, and management of datasets. * **Example:** Dataset creation, data inclusion, data usage transactions. As the Network evolves, more functionality and events will be added. Explore the sequencer below: [![Logo](https://explorer.eidon.ai/apple-touch-icon.png)Eidon Event Explorer](https://explorer.eidon.ai/) --- # Future Tokenomics | Eidon AI While the initial phases of the Eidon network utilize a points-based system, the long-term vision involves a transition to a token-based model. Upon the launch of the Eidon mainnet, the reward system will leverage the native EIDON token to directly compensate users for their data and other forms of participation. Upon launching our native Blockchain, we will recognize and reward the early contributors who helped build the network's foundation. The exact mechanics will be defined as the network approaches mainnet launch. The token-based reward system on mainnet will foster a sustainable ecosystem where users are incentivized to participate in various roles beyond data provisioning, including: * **Data Curation:** Contributing to the quality and organization of datasets. * **Data Validation:** Ensuring the accuracy and integrity of data within the network. * **Network Governance:** Participating in decision-making processes that shape the future of Eidon. * **Model Provisioning:** Submitting models to the network. * **Infrastructure Provisioning:** Provisioning critical infrastructure to power multiple network aspects. By aligning incentives with these essential functions, the Eidon network aims to create a robust and decentralized AI economy where all stakeholders are fairly compensated for their contributions. [PreviousEvaluating VLM MemAgents](/the-network/reward-mechanism/technical-details/evaluating-vlm-memagents) [NextSlashing Mechanism](/the-network/slashing-mechanism) Last updated 6 months ago --- # Slashing Reasons | Eidon AI [PreviousSlashing Mechanism](/the-network/slashing-mechanism) [NextStrikes and Bans](/the-network/slashing-mechanism/strikes-and-bans) Last updated 4 months ago Slashes are triggered by severe violations that undermine the network's integrity. The current violations that trigger slashings or bans are: * **Exact duplicate posts:** Submitting the exact same entry multiple times. * **User-Level duplicate posts:** Submitting multiple entries that are excessively similar to your own prior submissions. * **Network-Level duplicate posts:** Submitting multiple entries that are excessively similar to other user's submissions. * **Non-original content:** Uploading content sourced from external platforms, such as images from Google search results or clearly stock images/videos/audios. * **AI-generated posts:** Using generative AI tools to create data submissions without proper attribution or originality. Each violation has its own level of tolerance level and consequences outlined in the table in . [Slashing Mechanism](/the-network/slashing-mechanism) --- # Request Data | Eidon AI If you are a company or individual interested in requesting data, the Eidon network's Quest system offers a streamlined process for acquiring a variety of real-world, high-signal, ethically sourced multimodal data. We are open to discussing any requests for data and are happy to set up a conversation with your team. Please email `[[email protected]](/cdn-cgi/l/email-protection) ` with an overview of your data needs. [PreviousGet Rewarded](/get-involved/rewards-for-submitting-quests) [NextEidon FAQ](/get-involved/faq) Last updated 6 months ago ![Page cover image](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FoBbI1xgZVPSAURRMBSr0%252Feidon_4%2520%281%29.webp%3Falt%3Dmedia%26token%3D1f238887-f283-4460-a94c-94456c0dbb04&width=1248&dpr=4&quality=100&sign=a2cb3275&sv=2) --- # Sample Importance | Eidon AI [PreviousOutlier Likelihood](/the-network/reward-mechanism/technical-details/post-quality-memagent/outlier-likelihood) [NextPost Quality ( PQ ) Score](/the-network/reward-mechanism/technical-details/post-quality-memagent/post-quality-pq-score) Last updated 5 months ago While a low alignment score signals non-adherence to the instruction, a very high alignment score may indicate an _trivial_ sample, implying minimal challenge for the existing multi-modal AI MemAgents in the network. To ensure that novel and more challenging data receives higher rewards, we incorporate a generalized logistic function to assign an _importance weight_ to each sample: W(x,y)\=1−11+exp⁡(−k′(s(x,y)−τ′))\\mathcal{W}(\\mathbf{x}, \\mathbf{y}) = 1 - \\frac{1}{1 + \\exp(-k' (s(\\mathbf{x}, \\mathbf{y}) - \\tau'))}W(x,y)\=1−1+exp(−k′(s(x,y)−τ′))1​ where k′k'k′ controls the sensitivity to variations in the alignment score s(x,y)s(\\mathbf{x}, \\mathbf{y})s(x,y), and τ′\\tau'τ′ determines the inflection point at which the importance transitions from low to high. This function allows us to dynamically scale the reward based on the novelty and difficulty of each sample. By assigning higher rewards to novel and challenging examples, the model's training process prioritizes contributions that promote the network’s growth and adaptability. This adaptive weighting mechanism, enabled by the logistic function, ensures that the users are incentivized to submit data that meaningfully extends the network’s AI MemAgents. Importance Score ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FxpdtUAHySLPU83C0WFE8%252Fnovelty.png%3Falt%3Dmedia%26token%3D0830641c-bda5-405f-9031-e96f2b5e45bc&width=768&dpr=4&quality=100&sign=1d853a3d&sv=2) --- # Outlier Likelihood | Eidon AI [PreviousAlignment Score](/the-network/reward-mechanism/technical-details/post-quality-memagent/alignment-score) [NextSample Importance](/the-network/reward-mechanism/technical-details/post-quality-memagent/sample-importance) Last updated 5 months ago To detect out-of-distribution (OOD) samples in a multi-modal setting, we can leverage the alignment score s(x,y)s(\\mathbf{x}, \\mathbf{y})s(x,y) between a data point x\\mathbf{x}x and its corresponding instruction y\\mathbf{y}y. Since the Multi-modal alignment agents (e.g. Visual Language Models, Speech Language Models) are trained to embed semantically similar pairs close together, so a low alignment score reflects a divergence in semantic coherence between the two, indicating a possible OOD sample. We formalize this detection process within a probabilistic framework, defining the likelihood of a sample being OOD as follows: Lτ,k(s(x,y))\=1−11+exp⁡(−k(s(x,y)−τ))\\mathcal{L}\_{\\tau, k}(s(\\mathbf{x}, \\mathbf{y})) = 1 - \\frac{1}{1 + \\exp(-k(s(\\mathbf{x}, \\mathbf{y}) - \\tau))}Lτ,k​(s(x,y))\=1−1+exp(−k(s(x,y)−τ))1​ We then simply apply a threshold to assign an outlier flag: Iood(xi,xj)\={1if Lτ,k(s(x,y))<ϵo0otherwise\\mathbf{I\_{ood}}(\\mathbf{x}\_i, \\mathbf{x}\_j) = \\begin{cases} 1 & \\text{if } \\mathcal{L}\_{\\tau, k}(s(\\mathbf{x}, \\mathbf{y})) < \\epsilon\_o \\\\ 0 & \\text{otherwise} \\end{cases}Iood​(xi​,xj​)\={10​if Lτ,k​(s(x,y))<ϵo​otherwise​ In practice, threshold values τ,ϵo\\tau, \\epsilon\_oτ,ϵo​ are determined through precision-recall (PR) analysis on a development set, calibrating boundary control to maximize OOD detection accuracy. This approach enables precise, probabilistic separation of ID and OOD samples by aligning the boundaries with the semantic coherence objectives of the alignment model. Outlier Score ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FGAgvjy8A5of6ItunJRJZ%252Food.png%3Falt%3Dmedia%26token%3D68f69537-fdf0-41be-b938-26766f825645&width=768&dpr=4&quality=100&sign=1f7cd77c&sv=2) --- # Post Quality MemAgent | Eidon AI [PreviousTechnical Details](/the-network/reward-mechanism/technical-details) [NextAlignment Score](/the-network/reward-mechanism/technical-details/post-quality-memagent/alignment-score) Last updated 5 months ago The goal of the Post Quality (PQ) MemAgent is to understand (be able to describe) both the uploaded content x\\mathbf{x}x as well as the instruction y\\mathbf{y}y, compare them and assign a score 0≤s(x,y)≤10 \\leq s(\\mathbf{x}, \\mathbf{y}) \\leq 10≤s(x,y)≤1. This score reflects the degree of concordance between the content and its instruction, where higher values correspond to better alignment and, consequently, higher quality of the post. Furthermore, Open, DecAI landscape ( **The Wild** ) presents unique robustness challenges that demand resilience from the PQ MemAgent. Unlike controlled environments, the decentralized setting exposes the system to a myriad of unstructured, often unpredictable data sources. Content may be noisy, containing irrelevant or poorly structured information that blurs the intended meaning. Worse, some content is explicitly adversarial, crafted to deceive or exploit the system’s reward mechanisms. Thus, in this context, the PQ MemAgent’s resilience becomes paramount. It must distinguish between authentic and synthetic inputs, navigate ambiguous or misleading information, and accurately interpret content-instruction alignment even under noisy conditions. In the rest of this section -- we describe these key ideas in more detail. --- # Eidon State | Eidon AI [PreviousEidon FAQ](/get-involved/faq) [NextTransactions & Events](/the-network/transactions-and-events) Last updated 6 months ago The Eidon network is a state transition system, with the network state at any given point representing the canonical status of contributions for the creation of next-generation AI models. Currently, the Eidon State is tracked with a combination of offchain , storage, and a The state currently tracks: * Data Ownership * Data quality and rewards * Data usage * User accounts * Deployed MemAgents * Account balances The data itself is currently stored in centralized object storage providers, while each data post has a corresponding metadata object that is visible in the sequencer and holds information about the data's: * Origin * Ownership * Quality * Usage history We use this information to keep a canonical record of data provenance and data contributions to the network. In the future, the Eidon State will extend to track Models and other forms of contributions beyond just data. As we continue to navigate our , we aim to decentralize every piece of the network - from onchain contribution tracking to decentralized storage. [](#state-transitions) **State Transitions:** -------------------------------------------------- Currently, state transitions occur via . In the future, state transitions will occur through on-chain transactions triggered by user actions or MemAgents. These events/transactions update the state by modifying existing objects or creating new ones, all actions being openly viewable now and, in the future, subject to validator consensus. **State Transition Examples:** * A user submits a new data post, creating a new metadata object, altering the state. * When data is used, the corresponding dataset's usage statistics are updated, which triggers reward accrual for the data provider, altering the state. * Every 12 hours, the user reputation metrics are updated by a MemAgent, altering the state. [MemAgents](/the-network/memagents) [centralized sequencer.](https://explorer.eidon.ai/) [decentralization path](/the-network/path-to-decentralization) [sequencer events](/the-network/transactions-and-events) --- # Evaluating VLM MemAgents | Eidon AI [PreviousReward Accrual](/the-network/reward-mechanism/technical-details/reward-accrual) [NextFuture Tokenomics](/the-network/reward-mechanism/future-tokenomics) Last updated 5 months ago Evaluation on the Network is designed to be community-driven, leveraging collective insights to strengthen model alignment and robustness. To this end, we structured the data collection and labeling process as an interactive set of tasks within the network, where community members participate in labeling data points and assessing model outputs. We set up specific labeling tasks to capture a wide range of model behaviors, creating both clean and adversarial samples for a thorough analysis. These labeling tasks were designed to be user-friendly and accessible, ensuring broad participation across the community. To guide contributors, we provided detailed labeling guidelines, clearly defining the criteria for classifying outputs as "aligned" or "misaligned." This ensured that labels were applied consistently and accurately, even when handling complex or ambiguous outputs. For quality control, we implemented a consensus-based approach, where multiple contributors reviewed each data point. In cases of disagreement, the final label was determined by taking the median of all contributions, fostering a reliable, community-driven labeling process. This approach also includes mechanisms to reward high-quality, consistent labeling, ensuring that contributors are incentivized to maintain high standards. Through these community-driven labeling tasks, we are able to create a robust real-world dataset that reflects a diverse range of perspectives, strengthening our ability to evaluate and improve model alignment across different contexts. [](#alignment-score-distributions) Alignment Score Distributions --------------------------------------------------------------------- Box Plot (Fitting Normals) Violin Plots (Fitting Normal) Table 1. OOD Evaluation of VLM Reward MemAgents. ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FhMbhkvPUKs4i3uyD0tTs%252FScreenshot%25202024-11-12%2520at%25204.48.16%25E2%2580%25AFAM.png%3Falt%3Dmedia%26token%3Da0dbb096-9b8e-4a0a-803e-711278a6bc8b&width=768&dpr=4&quality=100&sign=eb88a75f&sv=2) ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2Flh7-rt.googleusercontent.com%2Fdocsz%2FAD_4nXe4ATw-DCiTccOnW4CPlaogfVbogJIq0MqHQwGyTuPBRNTcXbRR3p6Z8atGhFBztrPVFHs7QZQFf1apvFVY7Lka3rcu_i7iKhVtX4i13poLKAjsHCdVJOArxDjZDhaAnmdd7m_x3resW6qD94O2TTMrJcjs%3Fkey%3Dh1Ssaqvmteua0nGkW6FqNA&width=768&dpr=4&quality=100&sign=e0052afb&sv=2) ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2Flh7-rt.googleusercontent.com%2Fdocsz%2FAD_4nXc_X6qyOCnurPO5Yz3fd7BxKKa6K0RatqTvFj-RxsXwL0qfZLM8JG47nAEjdNkF4OWaRXk-zqxuDShRFjEXfq553PFU7u_LZlUyU8yqXQ-kQ3QEJoJ0ro8znekHtBfcWJ1EqblwbOem1ggB6ChxMpNvD3yr%3Fkey%3Dh1Ssaqvmteua0nGkW6FqNA&width=768&dpr=4&quality=100&sign=a5e870a1&sv=2) --- # User Quality (UQ) MemAgent | Eidon AI [PreviousRobustness to Sybil Attacks](/the-network/reward-mechanism/technical-details/robustness-to-sybil-attacks) [NextReward Accrual](/the-network/reward-mechanism/technical-details/reward-accrual) Last updated 3 months ago In addition to evaluating the quality of individual user uploads, we incorporate a measure of user consistency in our quality assessment model. Specifically, we compute a User Quality (UQ) Score, representing the user’s consistency in producing high-quality uploads over time. The UQ Score is calculated as an attention-weighted sum of historical Post Quality (PQ) Scores, where the attention weights follow a generalized sigmoid kernel. Let s^(xt,yt)\\hat{s}(\\mathbf{x}\_t, \\mathbf{y}\_t)s^(xt​,yt​) represent the post quality score of upload (xt,yt)(\\mathbf{x}\_t, \\mathbf{y}\_t)(xt​,yt​) at time ttt over a time horizon TTT, and let w(t)\\mathbf{w}(t)w(t) denote the attention weight applied to s^(xt,yt)\\hat{s}(\\mathbf{x}\_t, \\mathbf{y}\_t)s^(xt​,yt​). The user quality score QQQ is given by: s(u)\=∑t\=1Ts^(xt,yt)⋅w(t)∑t\=1Tw(t)s(u)= \\frac{\\sum\_{t=1}^{T} \\hat{s}(\\mathbf{x}\_t, \\mathbf{y}\_t) \\cdot \\mathbf{w}(t)}{\\sum\_{t=1}^{T} \\mathbf{w}(t)}s(u)\=∑t\=1T​w(t)∑t\=1T​s^(xt​,yt​)⋅w(t)​ where uu u denotes the user and w(t)\\mathbf{w}(t)w(t) is derived from a generalized sigmoid function: w(t)\=11+exp⁡(−k′′(t−τ′′))\\mathbf{w}(t) = \\frac{1}{1 + \\exp(-k'' (t - \\tau''))}w(t)\=1+exp(−k′′(t−τ′′))1​ This method assigns higher weights to quality scores at specific points in time, influenced by the choice of parameters k′′k''k′′ and τ′′\\tau''τ′′, allowing for flexible prioritization within the user history. For instance, with a high k′′k''k′′ value and τ′′\\tau''τ′′ positioned towards the end of the time horizon, the model will emphasize recent posts more heavily, giving less weight to older contributions. This approach thus balances a user’s historical quality contributions with their recent behavior, providing a comprehensive, time-aware assessment of user quality that reflects both long-term consistency and current engagement. User Quality ( Time / Upload ) UQ distribution on a i.i.d subset of Users ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FebwfDVE7cgDajhOdD9CN%252Fuq_att.png%3Falt%3Dmedia%26token%3D94c8f34c-a1b5-4e44-8c61-9178f9ee9b21&width=768&dpr=4&quality=100&sign=8585a68a&sv=2) ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FuKkhQpnjQqVfnktXncpR%252Fuq_distribution.png%3Falt%3Dmedia%26token%3Ddd7e17d0-610c-4f7c-9344-704cf0dc97f0&width=768&dpr=4&quality=100&sign=e4889d6e&sv=2) --- # Reward Accrual | Eidon AI [PreviousUser Quality (UQ) MemAgent](/the-network/reward-mechanism/technical-details/user-quality-uq-memagent) [NextEvaluating VLM MemAgents](/the-network/reward-mechanism/technical-details/evaluating-vlm-memagents) Last updated 5 months ago The reward is finally a combination of , and factors in s. Following the notations set in the previous discussions, the reward for a post x\\mathbf{x}x in response to data is denoted as: The reward in DecAI is designed as a combination of Post Quality (PQ) Score, User Quality (UQ) Score, and an adjustment for potential Sybil attacks. Using the notations established in previous discussions, the reward for a post, given a response to a data instance, is defined as: R(x,y)\=I(x,xj∈X)⋅max⁡(w0,w1⋅s^(x,y)+w2⋅s(u))\\mathcal{R}(\\mathbf{x}, \\mathbf{y}) = \\mathbf{I}(\\mathbf{x}, \\mathbf{x\_j} \\in \\mathcal{X}) \\cdot \\max \\bigg( w\_0, w\_1 \\cdot \\hat{s}(\\mathbf{x}, \\mathbf{y}) + w\_2 \\cdot s(u) \\bigg) R(x,y)\=I(x,xj​∈X)⋅max(w0​,w1​⋅s^(x,y)+w2​⋅s(u)) where the terms are - * I(x,xj∈X)\\mathbf{I}(\\mathbf{x}, \\mathbf{x\_j} \\in \\mathcal{X})I(x,xj​∈X) is an indicator function that checks for duplicate content within the set of posts X\\mathcal{X}X. If x\\mathbf{x}x is found to be highly similar to an existing post xj∈X\\mathbf{x\_j} \\in \\mathcal{X}xj​∈X, the indicator evaluates to 0, effectively setting the reward to zero to prevent duplicate contributions from receiving rewards. Otherwise, it evaluates to 1, allowing the reward calculation to proceed. () * s^(x,y)\\hat{s}(\\mathbf{x}, \\mathbf{y})s^(x,y) denotes the , assessing the quality and relevance of the post x\\mathbf{x}x with respect to the data y\\mathbf{y}y. * s(u)s(u)s(u) represents the , evaluating the reliability and past contributions of the user uuu who submitted the post. * Finally, w0,w1,w\_0, w\_1,w0​,w1​, and w2w\_2w2​ are weighting factors that balance the influence of each component in the reward calculation. This formulation combines content quality, user reputation, and duplicate detection to ensure that rewards are allocated fairly and that Sybil attacks and low-quality contributions do not undermine the integrity of the system. [Post Quality (PQ) Score](/the-network/reward-mechanism/technical-details/post-quality-memagent/post-quality-pq-score) [User Quality (UQ) Score](/the-network/reward-mechanism/technical-details/user-quality-uq-memagent) [potential sybil attack](/the-network/reward-mechanism/technical-details/robustness-to-sybil-attacks) [potential sybil attack](/the-network/reward-mechanism/technical-details/robustness-to-sybil-attacks) [Post Quality (PQ) Score](/the-network/reward-mechanism/technical-details/post-quality-memagent) [User Quality (UQ) Score](/the-network/reward-mechanism/technical-details/user-quality-uq-memagent) --- # Alignment Score | Eidon AI [PreviousPost Quality MemAgent](/the-network/reward-mechanism/technical-details/post-quality-memagent) [NextOutlier Likelihood](/the-network/reward-mechanism/technical-details/post-quality-memagent/outlier-likelihood) Last updated 5 months ago Given a data point x∈X\\mathbf{x} \\in \\mathcal{X}x∈X uploaded in response to instruction y∈Y\\mathbf{y} \\in \\mathcal{Y}y∈Y, the alignment score s(x,y)s(\\mathbf{x}, \\mathbf{y})s(x,y) quantifies how well the response data point adheres to the requirements set by the instruction. s(x,y):X×Y→\[0,1\]s(\\mathbf{x}, \\mathbf{y}) : \\mathcal{X} \\times \\mathcal{Y} \\to \[0, 1\]s(x,y):X×Y→\[0,1\] For ease of exposition, we will first assume that the instructions are free-form text, while the data uploaded in response is one of image, audio, video, or text. Later, we also show how these building blocks are combined to support multi-modal instructions and data. At a high level, the idea is to map both the user post x\\mathbf{x}x and the instruction y\\mathbf{y}y to a joint embedding space that fosters the proximity of semantically similar samples and separation of dissimilar ones. Then we can simply compute the alignment score using a suitable distance measure D\\mathcal{D}D, defined over the joint embedding space. Specifically, we assume access to mapping functions (encoders): fwX(⋅):X→Hf\_w^{\\mathcal{X}}(\\cdot): \\mathcal{X} \\to \\mathcal{H}fwX​(⋅):X→H and fwY(⋅):Y→Hf\_w^{\\mathcal{Y}}(\\cdot): \\mathcal{Y} \\to \\mathcal{H}fwY​(⋅):Y→H, mapping data and instruction to a joint embedding space where H∈Rk\\mathcal{H} \\in \\mathbb{R}^kH∈Rk denotes a separable Hilbert Space equipped with inner product operation. Given an instruction y∈Y\\mathbf{y} \\in \\mathcal{Y}y∈Y and corresponding user-uploaded data x∈X\\mathbf{x} \\in \\mathcal{X}x∈X, we can compute the embedding vectors (representation in the joint embedding space): zx\=fwX(x),zy\=fwY(y)\\mathbf{z}\_{\\mathbf{x}} = f\_w^{\\mathcal{X}}(\\mathbf{x}), \\quad \\mathbf{z}\_{\\mathbf{y}} = f\_w^{\\mathcal{Y}}(\\mathbf{y})zx​\=fwX​(x),zy​\=fwY​(y) Once we have vector representations of the instruction and corresponding sample on the shared embedding space, we measure alignment as the cosine similarity (normalized inner product) with a Rectified Linear Unit (ReLU) activation. s(x,y)\=max⁡(0,1τzxTzy∥zx∥∥zy∥)s(\\mathbf{x}, \\mathbf{y}) = \\max \\left(0, \\frac{1}{\\tau} \\frac{\\mathbf{z}\_{\\mathbf{x}}^T \\mathbf{z}\_{\\mathbf{y}}}{\\|\\mathbf{z}\_{\\mathbf{x}}\\| \\|\\mathbf{z}\_{\\mathbf{y}}\\|}\\right)s(x,y)\=max(0,τ1​∥zx​∥∥zy​∥zxT​zy​​) Intuitively, we project the sample and instruction onto a hypersphere S1k−1\=z∈Rk:∥z∥\=1τ\\mathcal{S}^{k-1}\_1 = {\\mathbf{z} \\in \\mathbb{R}^k : \\|\\mathbf{z}\\|=\\frac{1}{\\tau}}S1k−1​\=z∈Rk:∥z∥\=τ1​ and compute the _angular distance_ between the semantic embeddings of the sample and instruction. Here, τ∈R+\\tau \\in \\mathbb{R}^+τ∈R+ is a hyper-parameter that balances the spread of the representations on S1k−1\\mathcal{S}^{k-1}\_1S1k−1​. --- # MVP Rationale | Eidon AI [PreviousReward Mechanism](/the-network/reward-mechanism) [NextPoints Grounding](/the-network/reward-mechanism/points-grounding) Last updated 6 months ago During the initial phases, Eidon will utilize a points-based system to track and reward user contributions. Points serve as a measure of a user's engagement and value creation within the Network. [](#key-actions-rewarded) **Key Actions Rewarded:** -------------------------------------------------------- We structured our points system to be aligned with the actions that will most benefit the network at this early stage. * **Data Posting:** Users will earn points for submitting data, with higher rewards for higher-quality and in-demand data categories and modalities. Completing specific data Quests aligned with the needs of proposers will yield the vast majority of rewards for users. * **Invites:** Users are incentivized to grow the network by inviting new members, earning points for successful referrals. * **Daily Activity:** Consistent engagement is encouraged through daily tasks and prompts, rewarding users with points and streak multipliers for continued participation. The streak multipliers, in turn, boost your future rewards. The categories above represent the bulk of points given in the early phases of the Eidon Network. Additionally, miscellaneous engagement - at the team's discretion - will also be rewarded. This includes valuable community contributions, good stewardship, governance ideas, UX feedback, bug reports, etc. [](#factors-influencing-rewards) **Factors Influencing Rewards:** ---------------------------------------------------------------------- * **Data Quality:** A quality assessment mechanism evaluates alignment with Quest requirements, novelty, and the overall value of submitted data. This directly influences the rewards earned for a given post. See for a further breakdown of how this quality calculation happens. * **Data Modality:** Different data types (text, image, video) have varying levels of demand and complexity, hence impacting their value. For a more detailed overview of the proportional difference in value across modalities, see . * **User Quality:** Consistent high-quality contributions and adherence to community guidelines contribute to a user's quality score, which is, in turn, a factor in their overall rewards calcultion. * **Streak Multipliers:** Engaging in daily activities and maintaining a streak enhances the points earned for subsequent actions, promoting regular participation. * **Quest Points:** Submitting data for challenging or high-value quests will accrue significantly higher points than "simpler" quests. This incentivizes users to contribute to data categories in high demand. [Mechanics](/the-network/reward-mechanism/technical-details) [Points Grounding](/the-network/reward-mechanism/points-grounding) --- # Violation Tolerance | Eidon AI [PreviousStrikes and Bans](/the-network/slashing-mechanism/strikes-and-bans) [NextDispute & Guardian System](/the-network/dispute-and-guardian-system) Last updated 4 months ago Each slashable offense has its own tolerance levels as follows: Offense Tolerance Note **Exact duplicate posts** 5 Leads to progressive bans after tolerance is reached **User-Level duplicates** 5 Violation counter resets upon slashing **Network-Level duplicates** 10 Violation counter resets upon slashing **Non-original content** 5 Violation counter resets upon slashing **AI-Generated content** 3 Violation counter resets upon slashing All of these stats are visible on the user profile page under **Violations.** Your stats will look something like this: [](#exact-duplicate-case) Exact Duplicate Case --------------------------------------------------- We have a very low tolerance for exact duplicates. This means uploading the exact same piece of data more than once either on a single account or across multiple accounts. Unlike other offenses, exact duplicate posts do not lead to slashing right after the tolerance is reached. Instead, further offenses lead directly to progressive time bans culminating in a permanent ban from the network. #### [](#example) Example Say a malicious user takes a photo of a brick and tries uploading it twice for different tasks. While the first image will be accepted, the second image will be rejected and the user's exact match violation count will increase by 1. If they continue to engage in this behavior - either with the same image or different images - their violation count will continue to increase until it reaches the limit. When the limit is reached, further offenses trigger the following cases: * **1st offense above tolerance:** 24-hour ban * **2d offense above tolerance:** 48-hour ban * **3rd offense above tolerance:** Permanent ban Sample Violations count ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252F5beE3CeCWfSlTIOFh2OX%252Fimage%2520%283%29.png%3Falt%3Dmedia%26token%3Df93119ed-e08c-47eb-8495-3fd2b18e6dca&width=768&dpr=4&quality=100&sign=72b3a4b7&sv=2) --- # Post Quality ( PQ ) Score | Eidon AI [PreviousSample Importance](/the-network/reward-mechanism/technical-details/post-quality-memagent/sample-importance) [NextRobustness to Sybil Attacks](/the-network/reward-mechanism/technical-details/robustness-to-sybil-attacks) Last updated 5 months ago To capture both the alignment quality of a data sample and its likelihood of being an outlier, we introduce a Modified Alignment Score, integrating these factors to reflect the data’s informativeness and alignment consistency. This score enhances model robustness by weighting samples that not only align with task objectives but also exhibit high novelty—offering insights into challenging or atypical instances in the data distribution. The Modified Alignment Score s^(x,y)\\hat{s}(\\mathbf{x}, \\mathbf{y})s^(x,y) is defined as follows: s^(x,y)\=Iood(x,y)⋅\[11+exp⁡(−k(s(x,y)−τ))−11+exp⁡(−k′(s(x,y)−τ′))\]s(x,y)\\hat{s}(\\mathbf{x}, \\mathbf{y}) =\\mathbf{I}\_{ood}(\\mathbf{x},\\mathbf{y}) \\cdot \\bigg\[\\frac{1}{1 + \\exp(-k(s(\\mathbf{x}, \\mathbf{y}) - \\tau))} - \\frac{1}{1 + \\exp(-k' (s(\\mathbf{x}, \\mathbf{y}) - \\tau'))} \\bigg\] s(\\mathbf{x}, \\mathbf{y})s^(x,y)\=Iood​(x,y)⋅\[1+exp(−k(s(x,y)−τ))1​−1+exp(−k′(s(x,y)−τ′))1​\]s(x,y) The first logistic term models the probability of a sample’s alignment score falling outside the in-distribution boundary, while the second term accounts for novelty by penalizing samples that remain within a more conservative alignment range. By scaling the alignment score, the Modified Alignment Score _prioritizes samples that are both semantically challenging and well-aligned with network objectives_. [](#comparison-with-clip-score) Comparison with CLIP Score --------------------------------------------------------------- The performance metrics further reinforce this advantage: the PQ Score consistently outperforms the standard CLIP Score across precision, recall, F1 score, and accuracy. These gains highlight the PQ Score’s ability to prioritize samples that are both semantically challenging and well-aligned with task goals, enhancing the model’s robustness in detecting out-of-distribution instances. By capturing both alignment quality and novelty, our scoring mechanism offers a more nuanced approach to sample selection, ideal for tasks requiring reliable detection of novel but relevant data points. Metric Precision Recall F1 Score Accuracy CLIP Score **99.94** 95.71 97.78 95.70 PQ Score **99.93** **96.71** **98.30** **96.67** This formulation builds on previously discussed metrics for and . To validate the efficacy of our proposed scoring mechanism, we compare the standard CLIP Score with our Post Quality Score (PQ Score) on the CIFAR-100 dataset, using the OpenAI ResNet-50 model trained using Contrastive Language Image Pre-training () paradigm. As illustrated in the density plots, the PQ Score achieves a clearer separation between inliers and outliers, with minimal overlap between the distributions. This improved separation indicates that the PQ Score more effectively identifies challenging, novel samples while preserving alignment with network objectives. [OOD Probability](/the-network/reward-mechanism/technical-details/post-quality-memagent/outlier-likelihood) [Sample Importance](/the-network/reward-mechanism/technical-details/post-quality-memagent/sample-importance) [CLIP](https://openai.com/index/clip/) **Post Quality (PQ) Score** CLIP Score (CIFAR 100) (Ours) PQ Score (CIFAR 100) CLIP Score (Ours) PQ Score ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FPWuaEiPfnicQp2EUcpQz%252Food_al_scaled.png%3Falt%3Dmedia%26token%3D2fc16231-1414-4b4e-8ed0-d84b5ad6fdf8&width=768&dpr=4&quality=100&sign=a2309061&sv=2) ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FzFuTmJNbeKXQeC6nYa6P%252FCLIP-Distribution-cifar100-blip2-itm-vit-g.png%3Falt%3Dmedia%26token%3Da4fc8037-8fd1-40f5-a1f1-bb658749a1b4&width=768&dpr=4&quality=100&sign=a2858164&sv=2) ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252Fli3Wb8HBnQVhGBPmRrU3%252Fdist-mod-blip2-itm-vit-g.png%3Falt%3Dmedia%26token%3D2cb9384d-3434-4e46-af0c-e1d04cd0e0f0&width=768&dpr=4&quality=100&sign=97a94ce&sv=2) ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252Fs6alxYUtbbctMQb8jsFE%252Fcfm-raw-blip2-itm-vit-g.png%3Falt%3Dmedia%26token%3D24f5b058-635c-445a-9fee-3c73e718288d&width=768&dpr=4&quality=100&sign=add93f7c&sv=2) ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252Fo6QJZbGSHzoV9TycvDfT%252Fcfm-mod-blip2-itm-vit-g.png%3Falt%3Dmedia%26token%3D0c94d05b-8db2-47dd-aac9-89c5f3702288&width=768&dpr=4&quality=100&sign=a4288c67&sv=2) --- # Slashing Mechanism | Eidon AI [PreviousFuture Tokenomics](/the-network/reward-mechanism/future-tokenomics) [NextSlashing Reasons](/the-network/slashing-mechanism/slashing-reasons) Last updated 4 months ago Maintaining high-quality and high-integrity data is crucial for the Network's health. To mitigate spam and malicious data, we have implemented a Slashing mechanism that leads to temporary account suspension in certain offenses and, if persistent, permanent account ban. Below is the violations table containing the offenses, tolerance levels, and consequences. Each offense is explained in more detail in . Offense Tolerance Consequence Exact Duplicate Post 5 Progressive Ban User-Level Duplicates 5 Slashing + 1 Strike Network-Level Duplicates 10 Slashing + 1 Strike Non-Original Posts 5 Slashing + 1 Stirke AI-Generated Posts 3 Slashing + 1 Strike The actions that lead to violations are detected in a hybrid system that encompasses detection and review by our . Humans and agents work together to ensure a robust and fair system to prevent abuse and ensure the long-term health of the network. Note that we **will not penalize** users for . The slashing system is built to prevent only abuse and dishonest behavior. We strongly encourage all of our community members to avoid the actions above and upload **HONEST DATA.** **We want to see your world, environment, and way of thinking.** If the on the app are not giving you enough options to upload honest data, please reach out to us on Twitter or Discord and help us ideate more Quests to empower you to collect data around you. To dispute the violations, slashes, or bans, please email [\[email protected\]](/cdn-cgi/l/email-protection) . [Slashing Reasons](/the-network/slashing-mechanism/slashing-reasons) [MemAgent](/the-network/memagents) [Guardian System](/the-network/dispute-and-guardian-system) [non-aligned or outlier data](/the-network/reward-mechanism/technical-details/post-quality-memagent/outlier-likelihood) [Quests](/get-involved/explore-quests) --- # Reward Mechanism | Eidon AI [PreviousMemAgents](/the-network/memagents) [NextMVP Rationale](/the-network/reward-mechanism/mvp-rationale) Last updated 5 months ago > _A cornerstone of this approach is a fair and transparent reward mechanism, which not only incentivize user participation but also promotes high-quality contributions essential for training robust AI models._ To this end, in this work we develop a **multi-agent dynamic point-based system** aimed at allocating rewards in proportion to the quality and relevance of user-generated content. This enables the users to receive rewards that reflect the practical value of their contributions across several dimensions, including content originality, alignment with AI objectives, data quality, and adherence to community standards. This framework allows for the flexible assignment of points based on the evolving needs of the decentralized network, providing a foundational mechanism for future token distributions and monetary rewards as the network matures. [](#core-principles) **Core Principles:** ---------------------------------------------- Eidon's reward system is rooted in the ideals of fair compensation and radical transparency. Users are rewarded for contributions that enhance the network's value. As we progress on our towards full trustlessness, we want to be radically transparent about every parameter that goes into calculating user rewards for data submissions — from underlying functions and models used to overarching rationale and guiding principles. We hope that this radical transparency can foster community-directed governance as the network evolves towards becoming fully decentralized. [decentralization path](/the-network/path-to-decentralization) --- # Decentralization Path | Eidon AI [PreviousDeveloper API](/the-network/developer-api) Last updated 6 months ago We want Eidon to become the foundational layer for open AI development in the future. We want a world where anyone can contribute to the development of AI, and where the technology is owned by the those who contributed to its creation. Our goal is for Eidon to become that canonical unbrokable record of contribution. But for that to work, that "record of contribution" must be decentralized. It can't be owned, controlled, or unilaterally altered by any single party. Every aspect of the Eidon Network must become a public good, owned by anyone who participates. This path will be non-linear, but the direction is clear: Full decentralization. And our philosophy as we walk that path is that of radical transparency. Below are the main phases and components of the network that will be decentralized. [](#phase-0) Phase 0 ------------------------- * Launch initial , bringing online to users. * Launch our to begin the canonical record keeping. * Incentivize individuals and companies to contribute diverse, high-quality data to the network. * Most elements of the network are at this stage centralized. [](#chain-launch) Chain Launch ----------------------------------- * Open our for permissionless integration, allowing users to create their own Eidon clients. * Foster a robust ecosystem of developers building innovative applications on the Eidon platform. * Expand the possible contributions type beyond data provisioning, opening opportunities for more people to participate in the network. * Decentralize the sequencer with the Eidon chain. * Launch our governance token. * Post Quality weights and functions become ruled by community governance. [](#storage) Storage ------------------------- * Decentralized the network data and model storage. * Build incentives for storage provides. [](#code) Code ------------------- * Open source all of our code * Open the development of the network to be carried out by the community in an open-source fashion. [](#model-infrastructure) Model Infrastructure --------------------------------------------------- * Build new node types on the network that perform multiple tasks such as end-user model inference or model creation. * Decentralize the Post Quality oracle. [Applications](/get-involved/post-data) [Quests](/get-involved/request-data) [open sequencer](/the-network/transactions-and-events) [API](/the-network/developer-api) --- # Dispute & Guardian System | Eidon AI [PreviousViolation Tolerance](/the-network/slashing-mechanism/violation-tolerance) [NextDeveloper API](/the-network/developer-api) Last updated 4 months ago > See with the eyes of wisdom, speak with the voice of compassion A core principle of Eidon is fairness. To ensure the reward mechanism remains just and accounts for potential misclassifications, we implemented a Dispute Resolution System through Community Guardian. This system allows users to challenge automated model decisions and provides a human oversight layer for content moderation and quality control. [](#disputing-posts) Disputing Posts ----------------------------------------- Users can dispute any of their posts flagged as outliers by the Post Quality (PQ) MemAgent. This is crucial for several reasons: * **Ensures Fairness:** Automated systems are not perfect. Disputes provide a recourse for users who believe their content has been misclassified. * **Model Improvement:** Disputes provide valuable feedback to refine our MemAgents, helping them learn and adapt to nuanced content. By highlighting edge cases and challenging misclassifications, the system continuously improves MemAgent's accuracy and understanding of the real world. * **Community Engagement:** The dispute process encourages users to actively participate in maintaining the quality and fairness of the reward system. If a post is successfully disputed, the user who made the post will earn 50% of the maximum points possible for that modality in that quest. For example, if the user successfully disputes a **video** post for a Quest where videos are worth 100 points, the user will earn 50 points after successful dispute resolution. Successful disputes are how the model learns. They are a fundamental part of the Eidon human-agent teaching loop and may earn additional rewards in the future. ### [](#how-to-dispute-a-post) How to Dispute a Post Disputing posts is a simple process that can be performed directly on the . 1. Go to your `Previous Posts`. 2. Select the **outlier** post you want to Dispute. 1. You can only dispute outlier posts. 2. Outlier posts are marked by ❗ on the app. 3. Click the big red dispute button. And that's it, your post is queued up for dispute and review by our Guardian System. ### [](#preventing-spam) Preventing Spam Disputes are essential for the health and fairness of the network. But users should use them responsibly. There are no limits to the number of outlier posts you can dispute, but unsuccessful disputes may decrease your user reputation. Further, excessive unsuccessful disputes may also lead to slashing and account banning. [](#guardian-system) Guardian System ----------------------------------------- The Guardian system provides human oversight for disputes and content that falls within the "confusion zone" of the Eidon Network alignment score. This zone represents content where the PostQuality MemAgent had low confidence in its classification (near the threshold τ). Guardians are the highest trusted community members responsible for reviewing flagged content and disputed posts, providing a crucial human element in the decision-making process. ### [](#becoming-a-guardian) **Becoming a Guardian** Guardians are selected from the pool of Acolytes, who in turn are community members who have demonstrated consistent high-quality contributions and adherence to community guidelines. The selection process for Guardians involves: 1. **Acolyte Status:** Applicants must first achieve Acolyte status through active and positive participation within the Eidon Network. 2. **Application and Interview:** Interested Acolytes apply and undergo a short interview process with the Eidon team. This interview assesses their understanding of the network's principles, the reward mechanism, and their commitment to fairness and accuracy. 3. **Good Standing Requirement:** Guardians are expected to maintain good standing within the network. Unaligned behavior, such as biased labeling or abuse of the Guardian role, will result in penalties, including point reductions up to and including full points slashing. ### [](#guardian-responsibilities-and-rewards) **Guardian Responsibilities and Rewards** Guardians are essential for maintaining the integrity and fairness of the Eidon Network. Their responsibilities include: * **Reviewing Disputed Posts:** Carefully assess user appeals for posts flagged as outliers, considering the context and instructions. * **Assessing Confusion Zone Content:** Evaluate content where the PQ MemAgent has low classification confidence. * **Providing Clear and Concise Feedback:** Explain the rationale behind their decisions to users, fostering transparency and understanding. Guardians are compensated for their time and effort with points awarded per labeling action. The reward structure aims to make it a viable option. The current reward range is ~1-5 points per labeling event, with the exact amount depending on the complexity and type of content being reviewed. This compensation model recognizes the valuable contribution Guardians make to the Eidon Network. Future development may include mechanisms for dynamic reward adjustment based on community needs and Guardian performance. [app](https://app.eidon.ai) ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FHAdvT3AcTRf4juSw77UZ%252FIMG_A8D76C22DB06-1.jpeg%3Falt%3Dmedia%26token%3D89765004-7cc7-41e9-b974-5b98916113fe&width=768&dpr=4&quality=100&sign=5f149b8c&sv=2) ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FixBIvNMyYpmKqv4N5fiW%252FIMG_DE002EB14D90-1.jpeg%3Falt%3Dmedia%26token%3Df8719b32-82e6-4acd-b7dd-cdff6d8c21d3&width=768&dpr=4&quality=100&sign=42c0ca42&sv=2) --- # Get Rewarded | Eidon AI [PreviousPost Data](/get-involved/post-data) [NextRequest Data](/get-involved/request-data) Last updated 6 months ago You will earn points for every data post submitted. Points are a measure of contributions to the Eidon network during its early phases. Each Quest contains the **hypothetical maximum amount** of points a **single** post for that Quest can earn. The exact reward for a given post will depend on multiple factors such as the modality, the user quality score, post quality score, and streak multiplier, among other parameters. As seen above, different modality types for the same Quest can earn different points. Generally, video will be the most valuable form of data, followed by audio, followed by images. Similarly, certain categories of data will be more valuable than others (for instance, first-person POV tasks will be worth more than coffee shop interior data). Overall, our reward mechanism is designed to **incentivize high-quality data** and penalize spam. You can see a , including all the parameters that may influence your rewards. [full description of our reward mechanism here](/the-network/reward-mechanism) ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FWGTExL4ihAfzTSSFEMWA%252FScreenshot%25202024-10-01%2520at%252010.03.03%25E2%2580%25AFAM.png%3Falt%3Dmedia%26token%3D7c25fb4a-1e48-4d3c-870f-458d9bfdb048&width=300&dpr=4&quality=100&sign=a5815a7a&sv=2) ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FQanJlOtLyCzSp8p96Yyu%252FScreenshot%25202024-10-01%2520at%252010.00.25%25E2%2580%25AFAM.png%3Falt%3Dmedia%26token%3D5cb16f74-096d-4b6d-9595-9169405503e5&width=300&dpr=4&quality=100&sign=d39c129c&sv=2) ![Page cover image](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FESvsXWYA53Jcj0naGlUO%252Feidon_9.webp%3Falt%3Dmedia%26token%3D6b8d80a3-4421-4715-868f-21970164aab5&width=1248&dpr=4&quality=100&sign=cae808f8&sv=2) --- # Eidon FAQ | Eidon AI [PreviousRequest Data](/get-involved/request-data) [NextEidon State](/the-network/eidon-state) Last updated 4 months ago [](#general) General ------------------------- #### [](#is-there-a-token) **Is there a token?** No, there is no live token for the Eidon Network. Be mindful of scams and always refer to the official channels and links. #### [](#what-are-eidon-points) What are Eidon Points? The initial phases of the Eidon network utilize a points-based system, but we will eventually transition to an on-chain token-based model. Eidon points are our way to quantify contributions in the early stages of the network, which will be recognized later on. Official announcements will be made when we transition to a blockchain-based system. For more detailed information about tokenomics, please refer to the . [](#participation) Participation ------------------------------------- #### [](#how-to-become-an-acolyte) How to Become an Acolyte? To become an Acolyte, you'll need to join our Discord. Once invited, you must create an introduction post in the _#introductions_ channel of our Discord server. This helps us get to know you and ensures you're aligned with our community values. Please follow the instructions provided in _#introductions_ carefully! #### [](#how-to-become-a-guardian) How to Become a Guardian? Guardians are trusted community members responsible for reviewing disputed content and maintaining the integrity of the Eidon network. They are selected from active Acolytes who have demonstrated high-quality contributions and adherence to community guidelines. The process involves 1. An application 2. A short interview 3. Having and maintaining good standing within the network. [](#collecting-data) Collecting Data ----------------------------------------- #### [](#any-considerations-when-collecting-data) Any considerations when collecting data? When collecting data in public spaces or collecting data from third parties, we ask that you: * Avoid capturing identifiable faces * Ensure consent from people other than you who directly appear on the data * Be mindful of local regulations or rules around data capture #### [](#how-to-dispute-posts) **How to Dispute Posts?** Disputing consists of reporting a mistake the Post Quality MemAgent made in grading one of your posts. Disputes are a fundamental part of the Eidon Network, enabling our agents to learn and evolve. When you dispute a post, you queue it up for review by our Guardian system. To dispute: 3. Click the big red `Dispute` button. #### [](#how-to-delete-a-post) How to delete a post? Similarly to disputing, you can delete by: 2. Select the post you want to delete. 3. Click the `Delete` button. Note that once you delete previous data, you will lose any previous or future rewards associated with it. Note also that if a deleted post has been downloaded off the network by a third party, deleting the data will prevent future access. However, it can't retroactively eliminate previous accesses. [](#geotagging-posts) **Geotagging Posts** ----------------------------------------------- Sharing your location with Eidon allows you to earn a 10% bonus on every geotagged post, and participate in Quests specific to a given area. #### [](#to-geotag-your-posts) **To Geotag your posts:** If you haven't already granted permission for location access, Eidon will ask. You'll see a popup: **"app.eidon.ai Would Like to Use Your Location"** * Click **Allow** to share your location and start earning your bonus! #### [](#managing-your-location-sharing) **Managing Your Location Sharing:** * **Stop sharing:** Visit the Settings page and toggle off Location Sharing to stop sharing your location. **Troubleshooting Location Sharing:** #### [](#on-your-user-profile-you-should-see-a-location-tab-with-your-current-location-permission-status) **On your user profile, you should see a Location tab with your current Location Permission Status** * `Granted` : You have granted permission * `Denied` : You have denied permission * `Prompt` : You have not yet granted or denied permission, and the browser is prompting the user to grant permission. On iOS, this status can also indicate that the location has been allowed but the Safari website settings are set to "Ask" periodically. * `Unsupported` : Browser does not support geolocation * `Disabled` : You turned Location Sharing off within the Eidon app settings page #### [](#on-ios) **On iOS:** * **Why does Eidon keep asking for my location?** This happens when Safari's location permission for app.eidon.ai is set to "Ask." Here's how to change it: 1. Open Safari on your iOS device. 2. Go to app.eidon.ai. 3. Tap the menu button on the left of the address bar, then tap "Website Settings." 4. Tap "Location" and change it to "Allow." * **Why doesn't Eidon ask for my location?** This might be because you previously selected "Don't Allow" or your browser's location services are turned off. Here's how to fix it: **Option A (Check Browser Permissions):** 1. Go to iOS Settings > Privacy & Security > Location Services. 2. Tap either "Safari Websites" or "Chrome" (depending on which you used to install Eidon). 3. Under "Allow Location Access," change it from "Never" to either "Ask Next Time Or When I Share." 4. If this doesn’t work, try Option B. **Option B (Check Safari Settings):** 1. Open Safari on your iOS device. 2. Go to app.eidon.ai. 3. Tap the menu button on the left of the address bar, then tap "Website Settings." 4. Tap "Location" and change it to "Allow." #### [](#on-android) **On Android:** * **Why does Eidon keep asking for my location?** This typically occurs when the browser's location permission for app.eidon.ai is set to "Ask every time." Here's how to change it (steps may vary slightly based on your device and browser): 1. Open your preferred browser on your Android device. 2. Go to app.eidon.ai. 3. Tap the three dots (menu) icon in the top right corner. 4. Tap "Site Settings" or "Permissions." 5. Tap "Location." 6. Change the setting to "Allow." * **Why doesn't Eidon ask for my location?** This might be because you previously selected "Deny" or your device’s location services are turned off. Here's how to fix it: **Option A (Check Browser Permissions):** 1. Open your preferred browser on your Android device. 2. Go to app.eidon.ai. 3. Tap the three dots (menu) icon in the top right corner. 4. Tap "Site Settings" or "Permissions." 5. Tap "Location." 6. Ensure the setting is either "Allow" or "Ask every time." **Option B (Check Device Location Settings):** 1. Go to your Android device's Settings. 2. Tap "Location." 3. Make sure "Location Services" are turned on. 4. Check the app permissions for your browser (e.g., Chrome) and ensure Location access is granted. Now you're all set to enjoy the benefits of location sharing with Eidon! [](#data-usage-and-privacy) Data Usage and Privacy ------------------------------------------------------- #### [](#who-owns-the-data-on-the-network) **Who owns the data on the network?** You do. You can always delete previously uploaded data while it is on the network. #### [](#how-does-eidon-use-the-posted-data) How does Eidon use the posted data? The data on the Eidon Network will be used to train the next generation of multimodal, embodied AI systems. This training may be performed openly on the network, or done off-network by a third party who purchases the data. Users will get compensated for the value their data generates at every step of the process. [](#still-have-questions) Still Have Questions? ---------------------------------------------------- We're always happy to help! \\ Guardians are compensated for their efforts. For detailed information about the Guardian system, responsibilities, rewards, and the selection process, please refer to the . On the , navigate to your previous posts. Select the post that you believe was wrongly marked as an . Read more about the . On the , navigate to your previous posts. Please refer to our for more details. If you have further questions, please don't hesitate to reach out through our official channels on or . [Future Tokenomics Documentation](https://docs.eidon.ai/the-network/reward-mechanism/future-tokenomics) [Guardian System Documentation](https://docs.eidon.ai/the-network/dispute-and-guardian-system) [app](https://app.eidon.ai) [outlier](/the-network/reward-mechanism/technical-details/post-quality-memagent/outlier-likelihood) [Dispute and guardian system here](/the-network/dispute-and-guardian-system) [app](https://app.eidon.ai) [Privacy Policy](https://github.com/Eidon-AI/eidon-disclaimer/blob/main/DISCLAIMER.md) [Discord](https://community.eidon.ai) [Twitter](https://x.com/eidon_ai) --- # Strikes and Bans | Eidon AI [PreviousSlashing Reasons](/the-network/slashing-mechanism/slashing-reasons) [NextViolation Tolerance](/the-network/slashing-mechanism/violation-tolerance) Last updated 4 months ago [](#slashing-strikes) Slashing Strikes ------------------------------------------- Getting slashed means having slashed by 1 point and receiving one strike. After any slashing, the violation tolerance for the offense that triggered the slashing event will be set back to zero; but the strike count will increase by one. If the user continues to misbehave and engage in dishonest behaviors, further slashes will lead to temporary bans and eventual permanent bans. It follows this rule: * **1st Strike:** Warning * **2nd Strike:** 24-hour ban * **3rd Strike:** 48-hour ban * **4th Strike:** Permanent ban The exception to this slashing strike system is the **Exact Duplicate Post** offense. This is covered in more detail in . #### [](#example) Example Say a user with user-quality `5` makes 5 non-original posts that are detected by the guardian system. At this point, the user has blown its tolerance for non-original posts. The user proceeds to upload another non-original post. This would lead to the user's first slashing event. Their user quality drops to `4`. The next time they log into the app, they will see a warning modal like this: At this point, the user is given another chance to correct their behavior. The non-original violation count resets to `0` and the user may continue to use the app. Say the user continues to engage in dishonest behavior by uploading 4 (one more than tolerated) AI-generated images. This would lead to the user's second slashing event. At this stage, the user will be locked out of their account for 24 hours. When they log back in, their user quality drops by one more point to `3` , their violation count for AI-generated content resets, and the user is given yet another chance to correct their behavior. If bad behavior continues and results in further slashing, the user will be banned for 48 hours at the third offense, and permanently banned thereafter. [](#ban-notice) Ban Notice ------------------------------- If you are banned, when you attempt to log into your account you will see a modal explaining to you the reason for banning and the time until your ban is lifted. It will look something like this: [user quality](/the-network/reward-mechanism/technical-details/user-quality-uq-memagent) [Violation Tolerance](/the-network/slashing-mechanism/violation-tolerance) ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FWg1GWyKNQBPAvELcH3r4%252Fimage.png%3Falt%3Dmedia%26token%3Da0b47a61-7c13-4ee8-a972-ec80f7773529&width=768&dpr=4&quality=100&sign=4a0a840c&sv=2) ![](https://docs.eidon.ai/~gitbook/image?url=https%3A%2F%2F1620916093-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FhdnizGODxOgjJ7dbsgvp%252Fuploads%252FP9ToGZUOrY9qqKD0zauO%252Fimage%2520%284%29.png%3Falt%3Dmedia%26token%3D967f175a-bb05-4af0-a91f-77398ed48773&width=768&dpr=4&quality=100&sign=5175367f&sv=2) --- # Robustness to Sybil Attacks | Eidon AI [PreviousPost Quality ( PQ ) Score](/the-network/reward-mechanism/technical-details/post-quality-memagent/post-quality-pq-score) [NextUser Quality (UQ) MemAgent](/the-network/reward-mechanism/technical-details/user-quality-uq-memagent) Last updated 5 months ago In **The Wild**, one of the most pernicious threats is **Sybil attacks**, where adversaries create numerous fake identities or user accounts to flood the system with low-quality or misleading content. This allows them to manipulate the reward mechanism and extract disproportionate incentives, diverting rewards away from genuine, high-quality contributions. In particular, duplicate data plays a critical role in enabling Sybil attacks. Attackers often rely on repetitive or redundant data submissions to maximize their rewards while minimizing effort. These low-quality contributions distort the reward allocation process, reducing the incentives available for authentic, valuable input. Duplicate data not only compromises the system’s integrity but also risks misaligned incentives, ultimately diminishing the value of contributions and undermining the system’s purpose of rewarding genuine engagement. To ensure the robustness and alignment of the reward mechanism, mitigating the effect of Sybil attacks is essential. Implementing strategies to detect and down-rank duplicate or near-duplicate submissions will protect the system from being gamed by fake accounts and ensure that rewards are allocated to high-quality, original contributions. [](#md5-hashing) MD5 Hashing --------------------------------- To tackle duplicate data detection in Sybil defenses, we leverage the Message Digest Algorithm (MD5), a cryptographic function that maps input data to a 128-bit hash. h(x)\=MD5(x):{0,1}∗→{0,1}128h(\\mathbf{x}) = \\text{MD5}(\\mathbf{x}) : \\{0, 1\\}^\\ast \\to \\{0, 1\\}^{128}h(x)\=MD5(x):{0,1}∗→{0,1}128 MD5 operates by first padding the data x\\mathbf{x}x to a multiple of 512 bits, appending a single 1 bit followed by sufficient 0 bits and the original input’s length, ensuring that identical inputs yield identical hash outputs, even for varying input lengths. MD5 then initializes four 32-bit buffers A, B, C, D with predetermined constants. A\=0x67452301,B\=0xEFCDAB89A = 0x67452301, \\quad B = 0xEFCDAB89A\=0x67452301,B\=0xEFCDAB89 C\=0x98BADCFE,D\=0x10325476C = 0x98BADCFE, \\quad D = 0x10325476C\=0x98BADCFE,D\=0x10325476 Each 512-bit block of the input is processed over 64 rounds using non-linear functions and bit-wise operations, producing a final 128-bit hash h(x)h(\\mathbf{x})h(x) as a concatenation of A, B, C, D. This hash serves as a unique _fingerprint_ for content, enabling efficient duplicate detection. [](#perceptual-hashing) Perceptual Hashing ----------------------------------------------- To address cases where exact duplicate detection may miss slightly altered content (such as minor variations in data), we utilize _perceptual hashing_ (pHash), a type of locality-sensitive hashing. Unlike cryptographic hashing (e.g., MD5), perceptual hashing generates hashes based on perceptual attributes of content, enabling approximate matching for near-duplicates. This approach is particularly effective for mitigating Sybil attacks involving minor content variations. A low Hamming distance indicates high similarity, making perceptual hashing well-suited for near-duplicate detection. By applying a threshold on the Hamming distance, we can flag minimally altered images or audio clips as duplicates, thereby strengthening defenses against Sybil attacks. [](#embedding-similarity) Embedding Similarity --------------------------------------------------- h(x)\=A∥B∥C∥Dh(\\mathbf{x}) = A\\|B\\|C\\|Dh(x)\=A∥B∥C∥D I(xi,xj)\={1if h(xi)\=h(xj)0otherwise\\mathbf{I}(\\mathbf{x}\_i, \\mathbf{x}\_j) = \\begin{cases} 1 & \\text{if } h(\\mathbf{x}\_i) = h(\\mathbf{x}\_j) \\\\ 0 & \\text{otherwise} \\end{cases}I(xi​,xj​)\={10​if h(xi​)\=h(xj​)otherwise​ For images, the pHash process starts by resizing and converting the image sample to grayscale. A Discrete Cosine Transform (DCT) is then applied, transforming the image into the frequency domain. The low-frequency DCT coefficients Di,jD\_{i,j}Di,j​, which encapsulate the core structural details of the image, are selected from a compact block in the top-left region of the transformed matrix. A binary hash, p(I)p(\\mathbf{I})p(I), is then generated by comparing each coefficient to the mean of these selected values: pk(I)\={1,Di,j\>μ0,Di,j≤μp\_k(\\mathbf{I}) = \\begin{cases} 1, & D\_{i,j} > \\mu \\\\ 0, & D\_{i,j} \\leq \\mu \\end{cases}pk​(I)\={1,0,​Di,j​\>μDi,j​≤μ​ where μ\\muμ is the mean of the selected DCT coefficients, and pk(I)p\_k(\\mathbf{I})pk​(I) represents the kkk\-th bit of the binary hash for the image. Similarly, for audio, pHash begins by dividing the raw audio signal into short frames, transforming each frame to the frequency domain via a Fourier Transform to obtain features such as a mel-spectrogram. From the frequency-domain output, the low-frequency components Fi,jF\_{i,j}Fi,j​, which contain the dominant spectral patterns, are retained. A binary hash p(A)p(\\mathbf{A})p(A) is generated by comparing each component to the mean of the selected frequency values: pk(A)\={1,Fi,j\>ν0,Fi,j≤νp\_k(\\mathbf{A}) = \\begin{cases} 1, & F\_{i,j} > \\nu \\\\ 0, & F\_{i,j} \\leq \\nu \\end{cases}pk​(A)\={1,0,​Fi,j​\>νFi,j​≤ν​ where ν\\nuν represents the mean of the selected frequency components, and pk(A)p\_k(\\mathbf{A})pk​(A) denotes the kkk\-th bit in the binary hash for the audio sample. To determine the similarity between two perceptual hashes, p(xi),p(xj);∀;xi∈Xp(\\mathbf{x}\_i), p(\\mathbf{x}\_j) ; \\forall ; \\mathbf{x}\_i \\in \\mathcal{X}p(xi​),p(xj​);∀;xi​∈X of two samples, we compute the Hamming distance — the count of differing bits between the hashes: ΔH(p(xi),p(xj))\=∑k∣pk(Xi)−pk(Xj)∣\\Delta\_H(p(\\mathbf{x}\_i), p(\\mathbf{x}\_j)) = \\sum\_{k} \\bigg| p\_k(\\mathbf{X}\_i) - p\_k(\\mathbf{X}\_j)\\bigg|ΔH​(p(xi​),p(xj​))\=k∑​​pk​(Xi​)−pk​(Xj​)​ I(xi,xj)\={1,if  ΔH(xi,xj)≤ϵ0,otherwise\\mathbf{I}(\\mathbf{x}\_i, \\mathbf{x}\_j) = \\begin{cases} 1, & \\text{if} \\; \\Delta\_H(\\mathbf{x}\_i, \\mathbf{x}\_j) \\leq \\epsilon \\\\ 0, & \\text{otherwise} \\end{cases}I(xi​,xj​)\={1,0,​ifΔH​(xi​,xj​)≤ϵotherwise​ To detect duplicates, we calculate the cosine similarity over a shared embedding space, using multi-modal agents to embed data points across different modalities. Let sijs\_{ij}sij​ denote the cosine similarity score between two data points in a joint embedding space obtained via AI MemAgents. s(xi,xj)\=max⁡(0,1τzxiTzxj∥zxi∥∥zxj∥)s(\\mathbf{x\_i}, \\mathbf{x\_j}) = \\max \\left(0, \\frac{1}{\\tau} \\frac{\\mathbf{z}\_{\\mathbf{x\_i}}^T \\mathbf{z}\_{\\mathbf{x\_j}}}{\\|\\mathbf{z}\_{\\mathbf{x\_i}}\\| \\|\\mathbf{z}\_{\\mathbf{x\_j}}\\|}\\right)s(xi​,xj​)\=max(0,τ1​∥zxi​​∥∥zxj​​∥zxi​T​zxj​​​) zxi\=fwX(xi),zxj\=fwX(xj)\\mathbf{z}\_{\\mathbf{x\_i}} = f\_w^{\\mathcal{X}}(\\mathbf{x\_i}), \\quad \\mathbf{z}\_{\\mathbf{x\_j}} = f\_w^{\\mathcal{X}}(\\mathbf{x\_j})zxi​​\=fwX​(xi​),zxj​​\=fwX​(xj​) In this setup, pairs of data points with a cosine similarity score sijs\_{ij}sij​ exceeding the threshold τ\\tauτ are flagged as duplicates (indicated by 1), while pairs below the threshold are considered unique (indicated by 0). This method allows DecAI to down-rank near-duplicate submissions by identifying high similarity in the shared embedding space, thereby protecting the system against Sybil attacks and ensuring rewards prioritize original, high-quality contributions. I(xi,xj)\={1if s(xi,xj)\>τ0otherwise\\mathbf{I}(\\mathbf{x\_i},\\mathbf{x\_j}) = \\begin{cases} 1 & \\text{if } s(\\mathbf{x\_i},\\mathbf{x\_j})> \\tau \\\\ 0 & \\text{otherwise} \\end{cases}I(xi​,xj​)\={10​if s(xi​,xj​)\>τotherwise​ --- # Email Protection | Cloudflare Please enable cookies. Email Protection ================ You are unable to access this email address eidon.ai ---------------------------------------------------- The website from which you got to this page is protected by Cloudflare. Email addresses on that page have been hidden in order to keep them from being accessed by malicious bots. **You must enable Javascript in your browser in order to decode the e-mail address**. If you have a website and are interested in protecting it in a similar way, you can [sign up for Cloudflare](https://www.cloudflare.com/sign-up?utm_source=email_protection) . * [How does Cloudflare protect email addresses on website from spammers?](https://support.cloudflare.com/hc/en-us/articles/200170016-What-is-Email-Address-Obfuscation-) * [Can I sign up for Cloudflare?](https://support.cloudflare.com/hc/en-us/categories/200275218-Getting-Started) Cloudflare Ray ID: **931d82de3d4c437b** • Your IP: Click to reveal 54.237.218.47 • Performance & security by [Cloudflare](https://www.cloudflare.com/5xx-error-landing) ---