Bittensor Subnet 70:Vericore

Evidence-Based Fact-Checking: Vericore is designed to process natural language statements and understand their meaning, then retrieve precise supporting evidence. Instead of a simple true/false, it returns quotes or data from credible sources to justify the conclusion​. Every answer from the subnet is accompanied by traceable source attributions for transparency​.

Dual Validation of Claims: The system provides both corroborating evidence and contradictory evidence (when available) for a given statement​. This dual output offers a balanced view of the claim’s factual landscape, highlighting what information supports the statement and what disputes it.

Semantic Intelligence at Scale: Vericore’s AI models perform deep semantic analysis on inputs, enabling them to interpret context and nuances of claims​. The subnet is scale-oriented – it incentivizes high-volume fact-checking operations, aiming to make comprehensive verification feasible at internet scale​. In essence, Vericore is built to handle a large throughput of fact-check queries without sacrificing accuracy.

Source Verification & Attribution: A core principle of Vericore is that all outputs must be verifiable. Miners don’t just answer whether a claim is true; they fetch relevant snippets from reliable sources (e.g. news articles, research papers, databases) to back up their assessment​. Every snippet includes source metadata or links, ensuring anyone can audit and follow up on the evidence provided​. This fosters trust and allows further investigation beyond the AI’s response.

These capabilities make Vericore a unique subnet in Bittensor: its “product” is trustworthy knowledge. The intended use cases include media fact-checking (e.g. verifying news or social media claims), academic and research validation, content integrity checks for publications, and any scenario requiring automatic source-backed verification of information​. By contributing this service, Vericore enriches the Bittensor ecosystem with a critical truth-verification layer.

How the Vericore Subnet Works (Technical Mechanics)

Subnet Architecture: In Bittensor, each subnet is an independent marketplace for a specific AI service. Vericore follows Bittensor’s general model of miners and validators working together within an incentive framework​. Miners are nodes that produce the AI service (in this case, fact-checking outputs), and validators are nodes that evaluate the miners’ work for quality and accuracy​. The subnet’s creator (dFusion in this case) provides an incentive mechanism – i.e. an off-chain code repository defining how miners and validators should perform their tasks and interact​. Vericore’s incentive mechanism is implemented in an open-source codebase (the Vericore GitHub repository) which specifies the interface and logic for fact-checking on Bittensor​.

Workflow: When a statement or claim is presented to Vericore (for example, via a validator’s API), the process works roughly as follows:

1. Query Intake: An input statement is received by the subnet’s interface. Vericore’s validator nodes run an API server to accept statements as input​ – this could be a user query or an internally generated test claim. For instance, a validator might expose a public endpoint where a client can submit a sentence like “The Eiffel Tower is over 300 meters tall.”.
2. Mining (Fact-Checking Task): The statement is broadcast to the Vericore miners. Each miner runs an AI model or pipeline tailored for fact-checking. They perform semantic analysis on the statement and gather evidence from various data sources. Miners may use large language models with retrieval capabilities (for example, querying search engines or a curated knowledge base) to find relevant information. The output from a miner is an evidence-backed validation of the claim – typically including one or more quoted passages and an assessment whether those passages support or refute the claim​. For example, a miner might return: “According to Wikipedia, the Eiffel Tower’s height is 330 meters​, which supports the statement.” Each miner’s response format is defined by the incentive code (ensuring consistency in what data they provide, such as source URL and a support/contradict flag).
3. Validation (Quality Evaluation): The Vericore validators independently evaluate the miners’ outputs. They run a suite of checks and models to measure the quality of each miner’s response. For Vericore, the validator’s code includes components like a domain validator, a quality model, and a snippet fetcher​.

• The domain validator may verify that the sources or content provided by miners are relevant to the query domain and not out-of-scope or malicious​. For instance, it ensures that the miner who answered about the Eiffel Tower actually provided evidence about the Eiffel Tower’s height (and not some unrelated or wrong context).
• The quality model quantitatively measures corroboration or refutation in the miner’s answer​. In other words, it assesses how well the evidence supports or contradicts the original statement and whether the miner correctly interpreted the information. This could involve an AI model that reads the quoted source text and the claim to judge alignment.
• The snippet fetcher is used to independently retrieve the referenced source material if needed​, ensuring miners aren’t fabricating quotes. The validator might double-check the snippet from its original source to confirm it was quoted correctly and in context.
• Additionally, Vericore’s validator has an active tester module (as seen in the code) which can generate its own test statements to continuously probe miners​. This helps ensure miners remain accurate even when not responding to external user queries – a sort of internal QA process.

Scoring and Consensus: Each validator then scores the performance of each miner over recent tasks, according to Vericore’s evaluation standards​. For example, a miner that consistently provides accurate, well-sourced evidence will earn high scores, whereas one that returns incorrect or low-quality evidence will score poorly. These scores are fed into Bittensor’s on-chain Yuma Consensus mechanism​. Yuma Consensus aggregates validators’ scores and determines the reward distribution (emission of TAO tokens and subnet incentives) for that epoch. Miners with better performance receive more rewards, incentivizing competition to provide the best fact-checking service​. Validators also earn rewards for their work in measuring miners, and the subnet creator can get a portion, aligning everyone’s incentives.

Result Delivery: The fact-checking result (supported by consensus of the best-performing miners) can be delivered to the requester. In practice, an application using Vericore might query several miners via a validator and then use the highest-quality response (or an aggregate of multiple) as the answer. Because all evidence includes sources, the end-user or application can verify the truth themselves.

This decentralized process is effectively a “distributed consensus mechanism for fact validation.” No single server or oracle decides what’s true; instead, many independent AI miners and validators collectively determine the reliability of information​. By combining advanced semantic AI with community-driven validation across multiple nodes, Vericore ensures that fact-checking is reliable, scalable, and transparent​.

Notably, Vericore’s design allows external integration: The presence of an API server on validators means that real-world applications can tap into the subnet’s service. For example, a news website could run a Vericore validator to feed its own content through for fact-checking, or an end-user could query Vericore via a community API to verify a claim they found online. (In Bittensor’s model, organizations often run validator nodes specifically to utilize the AI service for their applications​.) This opens the door for Vericore’s outputs to be used in products like browser fact-check extensions, social media truth filters, or research assistants, all while being backed by decentralized consensus rather than a single authority.

Subnet Parameters: On the blockchain side, Subnet-70 is identified by the token $SN70 (its “alpha” token representing stake in that subnet’s economy). Like other subnets, Vericore has its own staking and reward pool denominated in Bittensor’s TAO token. The subnet launched with a circulating supply of alpha tokens and a defined TAO market cap (e.g. on launch ~1,480 TAO in market cap)​. This economic structure allows TAO holders to delegate stake to Vericore’s miners/validators (if they believe this subnet will perform well) and earn a share of its rewards​. The Yuma validator (run by Yuma AI) provides institutional-grade validation and staking for subnets like Vericore, helping bootstrap their performance and security​. All of this ensures Vericore is not just a service, but a self-sustaining micro-economy within Bittensor, rewarding those who contribute to truthful AI.