Sentient Project Research Report: Building Loyal AI

Author: 0xjacobzhao Source: mirror

1. Project Introduction:

Sentient is an open source protocol platform dedicated to building a decentralized artificial intelligence economy. Its core goal is to establish an ownership structure for AI models, provide an on-chain calling mechanism, and build a composable and divisible AI Agent network. Through the **"OML" framework (Open, Monetizable, Loyal) and model fingerprint technology**, Sentient solves the fundamental problems of "unclear model ownership, untraceable calls, and unfair value distribution" in the current centralized LLM market.

The project is driven by the Sentient Foundation and focuses on the construction of open source AGI and protocol incentive mechanisms. The **"Loyal AI"** it advocates refers to an open AI model ecosystem that serves the community, is fairly governed, and can evolve over the long term.

Sentient Protocol's architecture consists of two core components: the blockchain system and the AI ​​Pipeline.

**AI Pipeline** is the basis for developing and training "loyal AI" artifacts, and includes two core processes:

-Data Curation: A community-driven data selection process for model alignment.

-Loyalty Training: A training process that ensures that the model remains consistent with the community's intentions.

The blockchain system provides transparency and decentralized control for the protocol, ensuring ownership and governance of AI artifacts. The main modules include:

-Governance: Controlled and decided by a decentralized autonomous organization (DAO).

- Ownership: Represent the ownership of AI artifacts through tokenization.

- Decentralized Finance (DeFi): Provide financial tools that support open, decentralized and fair governance and rewards.

2. Technical architecture and model confirmation mechanism:

1. OML model framework

In the white paper "Sentient: Loyal AI" (https://arxiv.org/abs/2411.03887) released in 2024, the project proposed the OML framework, which takes model confirmation as the starting point and attempts to build an "on-chain AI ownership protocol economy". For the first time, the concept of "AI native cryptography" was systematically proposed, aiming to provide an encryption-level ownership protection mechanism for open source models.

Interpretation of the core paper - "OML: Open, Monetizable, and Loyal AI":

**-Open: **The model must be open source, with transparent code and data structure, supporting community reproduction, auditing and forking;

-**Monetizable: **Each call of the model should trigger a revenue stream, which will be distributed to trainers, deployers and verifiers through on-chain contracts;

**-Loyal: **The model does not belong to the company, but to the contributor community, and the model upgrade direction and governance are determined by the DAO. Model ownership is verifiable, modification is restricted, and use is controlled

OML is not equivalent to a code license agreement, but through on-chain mechanisms and encryption methods, it ensures that the open source model has economic sovereignty and governance rights while maintaining openness. Construct an AI-native ownership and revenue protocol layer to ensure clear ownership, economic incentives, and behavioral governance even when the model is public.

Core Concept: AI-native Cryptography

Different from traditional cryptography that focuses on binary precision, AI-native cryptography uses the continuity, low-dimensional manifold structure, and differentiable characteristics of AI models to develop a lightweight security mechanism that is "verifiable but not removable." Its core technologies are:

Fingerprint embedding (Model Fingerprinting): insert a set of hidden query-response key-value pairs during training to form a unique signature of the model;

Ownership verification protocol: verify whether the fingerprint is retained in the form of a query through a third-party detector (Prover);

Permission call mechanism: before calling, it is necessary to obtain the "authorization certificate" issued by the model owner, and the system then authorizes the model to decode the input and return an accurate answer.

This method can achieve "behavior-based authorized call + ownership verification" without the cost of re-encryption.

Sentient currently uses Melange hybrid security: fingerprint confirmation, TEE execution, and on-chain contract profit sharing. The fingerprint method is the main line of OML 1.0 implementation, emphasizing the idea of ​​"optimistic security", that is, compliance by default, and detection and punishment after violations.

In addition, the paper defines a variety of verification mechanisms and attack models for the fingerprint mechanism: Fine-tuning attack (LoRA/SFT): modify some parameters to try to clear the fingerprint; knowledge distillation: build a student model to avoid the fingerprint of the parent model; parameter averaging (model merging): multi-model fusion attempts to dilute the fingerprint strength; fingerprint leakage attack: if the fingerprint key is leaked, the verification mechanism will fail.

The adversarial methods include: multiple rounds of embedding, subnet isolation training; combined benign samples to prevent forgetting; embedding weights to bind on-chain credentials.

OML and Sentient Protocol Protocol Architecture

The last chapter of the paper proposes a complete on-chain protocol (Sentient Protocol) to support OML:

**Storage layer:**Store model weights and fingerprint registration information;

**Distribution layer:**Authorization contract controls model call entry;

**Access layer:**Verify whether the user is authorized through proof of authority;

**Incentive layer:**The revenue routing contract distributes each call payment to the trainer, deployer and verifier.

It emphasizes compatibility with L1 main chains such as Ethereum and can be used as a standard protocol stack for on-chain AI ownership. All models are registered on the chain (including hash, signature, public key, fingerprint structure ID); Agent binds the registered model and triggers the submission of proof of use after each call; the contract distributes profits (trainer, deployer, verifier) ​​according to the call event and revenue flow rules; users or contract callers need to verify the call authority through the on-chain gateway. This structure constitutes the underlying design philosophy of Sentient Protocol.

2. Fingerprinting and model confirmation mechanism

GitHub: https://github.com/sentient-agi/oml-1.0-fingerprinting

This library is the first implementation of the Sentient fingerprint mechanism, providing a fingerprint injection and verification interface that can be embedded in the training process. Its purpose is to ensure that the model ownership can be verified, the usage behavior can be traced, and to prevent unauthorized copying and commercialization. This is a specific engineering implementation of the OML (Open, Monetizable, Loyal) framework.

**Model Fingerprinting Module (OML 1.0 Fingerprinting Module)** has the following functions:

**Behavioral Profiling:**Extract features from model behavior and generate unique fingerprints.

**Model Attribution:**Determine the source model based on the generated content.

**Fingerprint Verification:**Verify the source and consistency of the model output content.

The essence of the fingerprint mechanism is: by fine-tuning the model and embedding a unique set of "question-answer" (key-response) pairs, the model owner can verify whether the model belongs to him through specific queries, thereby forming a "cryptographic signature" of the model. In short, add a "watermark" to the LLM (large language model) to track and enforce the ownership of AI.

3. Enclave TEE computing framework

GitHub: https://github.com/sentient-agi/Sentient-Enclaves-Framework

Sentient Enclaves Framework is an open source framework that uses trusted execution environments (TEEs) such as AWS Nitro Enclaves to achieve secure deployment of model reasoning, fine-tuning, and proxy services. The framework emphasizes the "loyalty" of the model, ensuring that the model only responds to authorized requests and prevents unauthorized access and use. The security features are as follows:

**Payload encrypted deployment: **The model body, running configuration, input and output are all encrypted and stored in the enclave;

**TLS-based RPC: **Calls use a two-way TLS + attestation signature channel for communication;

**Log and cache isolation: **Run using the transient memory area and do not record permanent logs;

**Attestation report: **Each call generates an output certificate bound to the enclave signature to ensure that the execution behavior is traceable and verifiable.

TEE (Sentient Enclaves Framework) has the advantages of high performance and cloud integration, and is suitable for real-time AI and sensitive data processing, but is limited by hardware dependence and side channel attacks. Sentient Enclaves Framework uses Nitro Enclaves to provide enterprise-level privacy protection and a more friendly development experience. Compared with other encryption technologies, FHE provides strong privacy guarantees without hardware dependence and quantum security, but the performance overhead is huge, which is suitable for specific encryption computing scenarios and is difficult to directly replace TEE's high-performance tasks. ZK performs well in verifiability and decentralization scenarios and can be used as a supplement to TEE (this module is planned to be connected to zkML in the future).

4. Sentient Agent Framework

GitHub: https://github.com/sentient-agi/Sentient-Agent-Framework

Sentient-Agent-Framework is a lightweight open source framework that focuses on automating web tasks (such as searching and playing videos) by controlling browsers through AI agents, and provides a concise development experience (claimed to be 3 lines of code) by combining natural language instructions and LLM (such as OpenAI's GPT-4o). Its asynchronous execution, custom instructions, and multi-provider support make it suitable for rapid development and experimental applications. Combined with other projects of Sentient AGI (such as the TEE framework), it can be extended to security-sensitive scenarios. This architecture supports the construction of an intelligent agent with a complete closed loop of "perception-planning-execution-feedback", which can be expanded into an open source AI system with multi-agent collaboration, on-chain verification, and alignment.

Compared with traditional AI Agent Framework, Sentient-Agent-Framework has limited functions and is lightweight and concise. Compared with Crypto AI Frameworks such as Virtuals Protocol and ai16z (elizaOS), the latter provide diverse solutions in AI agent development, on-chain automation or Web3 integration, while Sentient-Agent-Framework is more suitable for off-chain Web tasks.

5. Sentient Social Agent

GitHub: https://github.com/sentient-agi/Sentient-Social-Agent

Sentient-Social-Agent is an AI system designed to build automated interactions on social platforms (Twitter, Discord, and Telegram), which can understand social environments, generate content, interact with users, and conduct social exchanges through multi-agent collaboration. It uses social perception, content generation, and behavior planning modules to support natural conversations and content creation on the platform. It is suitable for scenarios such as brand operations, virtual community management, and information diffusion. The system can be integrated with the Sentient Agent framework.

6. Open Deep Search (Not yet online)

On the Sentient official website, Open Deep Search is defined as a search agent that surpasses ChatGPT and Perplexity Pro. Team member Sewoong Oh disclosed part of the plan at the EthDenver 2025 Open AGI Summit:

Open Deep Search consists of two main parts: Sensient's search capabilities (including query restatement, URL and document processing, etc.) and reasoning agents. Reasoning agents use open source LLMs (such as Llama 3.1 and DeepSeek) to improve search quality through tools such as search, calculators, and self-reflection. On the Frames Benchmark, Open Deep Search outperforms other open source models and is even comparable to some closed source models, but since its functionality is not yet online, we are unable to evaluate its true capabilities.

III. Product form, implementation and planning

Currently, the products displayed on the Sentient official website are mainly the Sentient Chat chat dialogue platform and the open source model Dobby LLMs:

Sentient Chat: Sentient Chat is a decentralized AI chat platform launched by the Sentient Foundation, designed to provide a community-driven, customizable and loyal artificial intelligence experience. The platform combines open source large language models (such as the Dobby series) with an advanced reasoning agent framework, supports the integration of multiple tools, and meets the diverse needs of users. The core functions are:

1. **Open Reasoning Agent:**Sentient Chat's built-in reasoning agent is capable of performing complex tasks and supports the following functions: Search tool: integrated Open Deep Search (ODS) to provide real-time web search capabilities; calculator: handles mathematical calculations and data analysis tasks; code execution: generates and runs Python code to achieve complex logical reasoning and task execution.

2. **Multi-agent integration:**The platform supports the integration of multiple AI agents, and users can choose different agents to interact according to their needs, which improves the flexibility and functionality of the platform. Similar to the Web3 version of POE or the open, agent-driven Perplexity alternative

   Sentient Chat is currently in the beta stage and is only accessible through invitation codes distributed via email or community events. According to official information released to the public, more than 5,000 users have successfully obtained the right to use Sentient Chat, and more than 100,000 user queries have been processed. Since the author has not yet become a user of its test whitelist, it is currently impossible to evaluate the true capabilities of its model.

Dobby LLM Model Series: Fine-tuned based on Meta's Llama series, Hugging Face homepage (https://huggingface.co/SentientAGI)

1. Dobby-Unhinged Series Dobby-Unhinged-Llama-3.3-70B: Fine-tuned based on Llama 3.3-70B-Instruct, emphasizing personal freedom and cryptocurrency stance, with a straightforward, humorous and humane conversational style.

2. **Dobby-Mini-Unhinged-Llama-3.1-8B: **8B parameter version, suitable for resource-constrained devices, maintaining the core features of the "Unhinged" series.

3. Dobby-Leashed series: Dobby-Mini-Leashed-Llama-3.1-8B: Compared with the "Unhinged" version, it has a milder tone and is suitable for application scenarios that require more robust output.

Since the Dobby LLM model is a fine-tuned version based on Llama 3.1 and 3.3, we believe that its application scenarios are mainly in building chatbots, content generation and creation, role-playing agents, etc. Its advantages lie in flexible style generation, reasoning enhancement and low resource requirements, which are suitable for rapid deployment and flexible customization in resource-constrained environments. Compared with more powerful closed-source models such as GPT-4, Dobby LLM still has gaps in processing tasks involving advanced logic, cross-domain knowledge reasoning and deep reasoning.

IV. Ecosystem Cooperation and Implementation Scenarios

Currently, the Sentient Builder Program provides a grant of US$1 million to support developers in building AI Agents that run in the Sentient Chat ecosystem. Developers are required to use Sentient’s development kit and access its ecosystem through the Sentient Agent API.

At the same time, the ecological partners announced on Sentient’s official website cover project teams in multiple fields of Crypto AI (https://www.sentient.xyz/partners). The specific list is as follows:

**Model:**Eigenlayer, Move, CrunchDAO, Bagel, KGEN

**Intelligent agents:**Messari, Franklin Templeton, Kaito, MyShell, Third Web, Theoriq, Open, QNA3, Pond, Mira, Olas, Biconomy, Talus, Zettablocks, Axal, Morpheus AI, dFusion, ExponentAI, Fetch AI, Giza, JustTX, UnifAI, Questflow, QuillAI, Raiinmaker, Solo, Spectral, UOMI, PlayAI

**Data:**Kaito, Vana, The Graph, Space and Time, 0g, Open, QNA3, Zettablocks, Chainbased, dFusion, Dria/First Batch, Entrova, FractionAI, Hyve DA, Irys, Masa, Mizu, OpenLedger, Raiinmaker, Sapien, Zus Network

**Verifiable AI: **Nillion, Lagrange, pi2

**Blockchain: **Arbitrum, Polygon, Celo

**Infrastructure: **Lit Protocol, OpenGradient Sentient

As a leading project in the field of Crypto AI, its resource integration capabilities can cover any star startup project in the industry. However, it should be pointed out that the widespread existence of "marketing-type" cooperation in the Crypto field has created an illusion of false prosperity in the industry. The contribution and loyalty of Sentient's ecological partners to their ecology still require our continuous observation.

Open AGI Summit is a global conference organized by the Sentient team to explore the combination of artificial intelligence (AI) and encryption technology (Crypto). The author was fortunate to attend its summits during ETH Denver and ETHcc in 2024 and 2025. The Sentient team is able to gather the industry's top institutional investors and project entrepreneurs to participate, which is a highlight.

V. Team Structure and Research Background

Sentient Foundation brings together the world's top academic experts, crypto industry entrepreneurs and engineers, and is committed to building a community-driven, open source, and verifiable AGI platform. According to the official team information (https://sentient.foundation/people), its team members are mainly:

Core Leadership (Steering Committee)

Pramod Viswanath – Professor Forrest G. Hamrick of Princeton University, has long been studying information theory and communication systems, and leads Sentient's AI security and theoretical foundation construction.

Himanshu Tyagi – Professor at the Indian Institute of Science, specializes in privacy protection and decentralized learning algorithms, and provides academic support for model training and privacy collaboration.

Sandeep Nailwal – Co-founder of Polygon, responsible for blockchain strategy and global ecological layout, is a key figure connecting the crypto community and AI architecture.

Sensys Team – Web3 native product studio, leading user experience optimization and developer infrastructure construction, promoting the implementation of Sentient products.

**Core Engineering and Development Team:**Comes from well-known technology and blockchain companies such as Meta, Coinbase, Circle, Polygon, and Binance, as well as researchers from universities such as Princeton University, University of Washington, and Indian Institute of Technology. Team members have rich experience in LLM engineering, system security, computer vision, and data system construction.

**AI Research and Model Training Team:**The research team covers AI/ML, NLP, computer vision, and reinforcement learning, and its members have practical experience in institutions such as Google Research, Daimon Labs, and Fetch.ai. The composition of the team shows that Sentient has both strong academic depth and implementation capabilities and crypto ecological experience.

It should be pointed out that Sentient was founded with the aura of success of Sandeep Nailwal, the founder of Polygon. As an important expansion solution for the Ethereum ecosystem, Matic started with Plasma, a technology that is not leading but "cheap and fast" enough, to build a moat of differentiation for Polygon in NFT and social fields. At the same time, through the acquisition of Mir Protocol and Hermez Network and the launch of Polygon zkEVM, ZK technology was integrated into its blockchain expansion solution. As Sandeep Nailwal's second venture, Sentient's experience, funds, connections and market awareness are far more than those of the past. It can also raise huge funds in 2024 with its imperfect project concept, but the field of AI is different from Crypto after all. Sentient still faces external challenges such as changes in the new market environment, intensified competition and technological updates.

VI. Financing and Token Model

Financing time: mid-2024;

Financing amount: US$85 million (seed round);

Investors: Founders Fund, Pantera, and Framework Ventures jointly led the investment, and other VC institutions include Ethereal, Robot Ventures, Symbolic Capital, Dao5, Delphi, Primitive Ventures, Nomad, Hack VC, Arrington Capital, Hypersphere, IDG, Topology, Protagonist, Folius, Sky9, Canonical Crypto, Dispersion Capital, Mirana, Foresight, Hashkey, Spartan, Republic, Frontiers Capital, etc. (https://sentient.foundation/funding)

$SENT token usage (under planning) So far, Sentient has not launched an official token. Co-founder Sandeep Nailwal said that the project currently has no plans to issue tokens, but it may be evaluated in the future based on community needs and project development. Therefore, any project claiming to provide $SEN or other Sentient-related tokens should be treated with caution to prevent fraud.

According to the white paper, the possible uses of SENT tokens include: Agent incentive points are mapped to Tokens; used for model version management proposals and voting; staking to verify the authenticity of Agent output; DAO governance dividend mechanism, etc.

Sentient is a superstar project born with a golden key. Its investor background, financing scale and valuation are far beyond the reach of most Crypto AI projects on the market. On the one hand, its strong resource endorsement can more easily integrate various industry resources, and the high amount of financing can more easily hire top talents to join its team, and strong capital can support the development of the project through the industry cycle. On the other hand, the current Crypto industry is generally disenchanted with high-valuation projects endorsed by VCs. In addition, the price of VC coin projects is mainly based on capital operations and is seriously decoupled from fundamentals. Assuming that Sentient cannot deliver influential Crypto AI products and finally chooses to issue coins at a high valuation, it will also hurt the Crypto community that is in urgent need of rebuilding trust. How the team copes with the current industry dilemma is worth our continued observation.

Seventh, Competitive Product Analysis and Market Position

Most Crypto AI projects on the market focus on a single field such as data, models, computing, training or reasoning, or develop consumer-level applications such as AI Agent. Projects positioned as AI Chain include the AI ​​transformation of old blockchains (Near and ICP) or decentralized resource sharing coordination and token incentive protocols such as Bittensor, and Sentient's positioning does not fully match them. On the model training side, Sentient is more like an integration platform, and has a cooperative relationship with AI open source models on the market. On the Agent side, Sentient has some overlapping competition with Talus, Olas or Theoriq in terms of multi-agent systems and reasoning capabilities, but each project has different core goals and application scenarios and is still complementary.

VIII. Summary

As a decentralized artificial intelligence (AGI) protocol platform, Sentient aims to provide a clear ownership structure for AI models, and to call and distribute value through on-chain mechanisms to solve the unclear ownership and unfairness in the current centralized LLM market. The core framework OML (Open, Monetizable, Loyal) ensures the ownership, transparency and fair profit sharing of open source models through model fingerprints and blockchain technology. With the resources of Polygon co-founder Sandeep Nailwal, Sentient has gained support from many top VCs and AI ecosystem partners. Despite the uncertainty, controversy and competition in development, it is still expected that Sentient will become one of the standard protocols for decentralized AI ownership and promote the decentralized development of AGI.