The core challenge for AI in Web3 isn’t just proving that data hasn’t been tampered with; it’s proving the data’s substance. A cryptographic hash can guarantee a dataset’s integrity, but it cannot verify its origin, quality, or content. A creator could claim their AI is trained on financial data when it was actually trained on irrelevant information, and the hash would still be valid.
This “semantic gap” creates a trust vacuum, preventing developers from confidently building high-stakes applications that rely on third-party AI agents.
The Solution
Haithe is a trust protocol that solves this problem by creating a transparent, verifiable ecosystem for AI. We move beyond simple integrity checks to a system of substantive verification, enabled by three key roles: Creators, Consumers, and Auditors.
Instead of just registering a data hash, creators submit a detailed Claim Manifest about their dataset. Staked domain experts, acting as Auditors, are granted temporary, private access to investigate these claims. Their findings are published in a granular, on-chain Audit Report, allowing consumers to make informed decisions based on expert, independent verification.
How It Works: A Three-Part Ecosystem
Haithe’s workflow ensures checks and balances between all participants, with smart contracts on Hyperion orchestrating the process.
1. The Creator: Defining the Claims
Build & Register: Developers use the Alith framework to build high-performance AI agents and register them on the Haithe marketplace.
Submit the Claim Manifest: Alongside the agent, the Creator submits a structured manifest containing specific, falsifiable claims about the training data. For example:
Source: “API data from Uniswap v3 on Polygon.”
Date Range: “All swaps from June 1, 2023, to December 31, 2023.”
Content: “Dataset contains no personally identifiable information (PII).”
Schema: “Includes columns for token_pair, amount_in, amount_out, and gas_price.”
2. The Auditor: Investigating the Substance
Stake & Qualify: Domain experts (in finance, NLP, security, etc.) stake tokens to become eligible Auditors.
Accept the Task: An Auditor accepts an audit request, which can be funded by the Creator or a Consumer.
Investigate Claims: The Auditor is granted secure, time-limited, off-chain access to the private dataset. Their role is to meticulously investigate the claims made in the manifest. This requires genuine domain expertise.
Publish the Audit Report: The Auditor submits a detailed report to IPFS, with its hash recorded on-chain. This report confirms or refutes each claim individually, providing qualitative remarks.
3. The Consumer: Making Informed Decisions
Discover & Filter: Developers browse the Haithe marketplace for AI agents. They can filter for agents based on their audit status.
Review Reports: Before subscribing, a Consumer can review the detailed Audit Report for any agent. This allows them to assess if the agent’s verified characteristics fit their specific use case.
Integrate with Confidence: Armed with expert-verified information, the Consumer can confidently integrate the agent into their dApp, knowing exactly what they are getting.
Key Features
On-Chain Claim Manifests: Creators make specific, verifiable claims about their AI’s data foundation.
Expert-Led Audits: Audits are performed by staked domain experts, not just automated checkers.
Granular Audit Reports: Replaces a simple pass/fail with a detailed report that verifies individual claims.
Decentralized Access Control: A secure mechanism grants auditors temporary, private access to datasets for verification.
Incentive-Aligned Ecosystem: Creators are rewarded for building great agents, and Auditors are rewarded for providing high-quality, trustworthy verification.
Built for Performance: Powered by the Alith AI framework and the Hyperion blockchain for high throughput and low fees.
Technology Stack
Blockchain: Hyperion
AI Framework: Alith
Smart Contracts: Solidity
Decentralized Storage: IPFS / Arweave for datasets and audit reports
Community & Testing
Haithe is an ecosystem, and its strength comes from its community. We invite you to participate in our testnet by taking on one of the core roles:
Become a Creator: Build a simple agent with Alith, deploy it on our platform, and create a Claim Manifest for it.
Act as a Consumer: Browse the marketplace, review the claims and audit reports of existing agents, and integrate them into a test application.
Join the Auditor Program: If you have domain expertise, apply to our Auditor program to help verify claims and build the foundation of trust for the ecosystem.
Thanks for sharing this deep and original approach-Haithe sounds truly exciting! As a consumer, how can I be sure that an AI agent’s data truly comes from the sources claimed? Do the audit reports go into technical detail, and does relying on them require multiple auditors’ perspectives for confidence?
GM GM, Apologies for the delayed response , I was in the middle of relocating and just got everything settled. Thanks again for your interest in our project and for the thoughtful questions. Let me address each of them below
1) If the dataset is private, how do you ensure Auditors don’t misuse or leak it?
Great question. In our current model, auditors must undergo two layers of validation before they’re granted access:
First, they are verified by us (the platform) based on identity and domain expertise.
Second, they are validated by the DAO, ensuring community-level accountability.
While trust in auditors is still required to some extent, just like in real-world auditing , we are actively exploring ways to make this process more trustless.
2) Are Creators allowed to challenge an audit report if they think it’s unfair or inaccurate?
Yes, it’s a crucial safeguard.
We agree that auditors shouldn’t have unchecked authority. Creators will have the ability to challenge audit reports, either by:
Requesting a re-audit (potentially by a different auditor or group of auditors), or
Submitting a formal rebuttal that gets appended to the audit report for transparency.
3) What kinds of use cases are you targeting first – trading bots, oracles, LLMs?
Our initial focus is on trading bots and LLMs. These use cases have both high impact and urgent need for trust and verifiability.
Helloo, Apologies for the delayed response, I really appreciate your thoughtful question!
As a consumer, audits are exactly what enable trust that an AI agent’s data truly comes from the sources claimed. Each audit report goes into technical detail, verifying every claim made in the agent’s Claim Manifest — like the data source, date range, schema, and more.
And yes, confidence naturally increases when multiple auditors review the same dataset. The more popular or high-quality a dataset is, the more likely it is to attract repeat audits, giving you a broader and more trustworthy view.
Thanks for the detailed response, and hope the move went smoothly!
I really like the dual-layer validation approach and the transparency mechanisms for audit disputes.
Haithe seems to tackle a real pain point, transparency in AI data verification. Love the Claim Manifests and real expert Auditors putting skin in the game. This goes way beyond simple hashes, giving builders and consumers real trust and clarity. Excited to see how it grows!
Hey Everyone, We’re doing a community research to understand how developers, data scientists, auditors, and Web3 builders think about trust, verification, and integrating third-party AI.
Please take 3–5 minutes to fill out this quick questionnaire
Your input will directly shape Haithe’s early design and testnet launch.
Basic Info
What’s your background? (Select all that apply)
AI/ML Developer
Web3 Developer
Data Engineer / Analyst
Security Researcher
Domain Expert (finance, law, etc)
Founder / Builder
Student / learner
Other technical background
Other non-technical background
0voters
How familiar are you with verifiable data or trust protocols?
Not familiar
Somewhat familiar
Very familiar / already using similar tools
0voters
Have you ever used an LLM created by someone else (e.g., from HuggingFace, GitHub, or some DAO)?
Yes, Frequently
Occasionally
Rarely
Never
Not sure
0voters
What are your biggest concerns when using external AI models? (Select up to 3)
Not knowing the source of training data
Potential misuse of personal or sensitive data
Legal or ethical issues
Poor model performance
Incompatibility with my stack
Lack of transparency in how the model was built
No concerns
0voters
What would most increase your trust in using someone else’s AI model? (Select up to 3)
Transparent training data claims
Independent third-party audits
Verifiable source data (e.g., on-chain or timestamped)
Reputation of the creator
Community reviews or ratings
Open-source code and training pipeline
Nothing — I already trust most open models
Nothing — I don’t trust external AI models
0voters
How valuable do you think it is to have expert audits of AI models’ data claims?
Not valuable
Slightly useful
Useful
Very valuable
Absolutely essential for critical applications
0voters
If you had to choose a role in the Haithe ecosystem, which would you most likely want to try?
Creator (build and publish AI agents)
Auditor (verify dataset claims)
Consumer (use verified agents in dApps)
None right now — just observing
0voters
Would you stake tokens or reputation to act as an Auditor in return for rewards?
I went through your proposal and i have some questions to ask:
How does the “Rust-based optimization” work for inference, and what kind of performance gains can be expected?
Given that Hyperion is a Layer 2 solution for AI execution, what are its specific advantages over other Layer 2s for this use case?
The proposal mentions “secure, time-limited, off-chain access to the private dataset.” Can you elaborate on the specific technical mechanisms for this? Is it through Zero-Knowledge Proofs (ZKPs), Trusted Execution Environments (TEEs), federated learning, or another method?
While the hash of the report is on-chain, how can consumers be confident in the truthfulness and completeness of the qualitative remarks within the report, beyond just its integrity?
Please take 3–5 minutes to fill out this quick questionnaire
