This episode reveals how Mantis is engineering a Sybil-resistant prediction market on Bittensor, using a novel incentive mechanism to distill high-value financial signals from a decentralized network of miners.
The Unlikely Origin Story
- Barbarian shares his unconventional entry into the crypto world. While living in an isolated part of Costa Rica with no home internet, his family discovered Bittensor through a chance encounter at a local café.
- Initially drawn to the profitability of mining, Barbarian started with no coding or command-line experience.
- At the time, Bittensor documentation was sparse, assuming a high level of technical expertise among its users.
- This forced a steep learning curve, as he navigated the ecosystem without modern tools like ChatGPT. Barbarian notes, "At that point I actually did not know how to code nor even how to operate a computer via a command line."
From Novice Miner to Subnet Dominance
- Barbarian details his rapid evolution from a beginner to a dominant force on Bittensor's Subnet 8, the original time-series prediction network. This experience provided critical lessons that directly informed the design of Mantis.
- His success on Subnet 8, which focused on predicting Bitcoin's price eight hours in the future, laid the groundwork for Mantis.
- He observed firsthand the economic flaws of early prediction subnets, where Sybil attacks—the practice of running many low-quality models to gain rewards through volume—were the most profitable strategy.
- This insight became the central problem Mantis was designed to solve: creating a system where genuine predictive accuracy is the only path to profitability.
Mantis: The Core Goal and Methodology
- Mantis aims to create a robust prediction market by selecting valuable time-series data and building a system that rewards genuine, high-quality signals while penalizing low-effort or duplicative contributions.
- The primary goal is to overcome the historical effectiveness of Sybil mining, where operators register a large volume of UIDs with models that lack long-term predictive power.
- The core methodology involves training network-wide ensemble models, which combine the outputs from all miners to create a single, more powerful prediction.
- By focusing on the collective intelligence of the network, Mantis can identify and reward miners who contribute unique and valuable information, rather than those who simply copy others or submit noise.
Technical Architecture: How Mantis Works
- Barbarian outlines the lean and efficient technical architecture of Mantis, emphasizing a design that prioritizes security, efficiency, and accurate signal attribution.
- Miner Communication: Miners upload time-locked encrypted embeddings to a public S3 bucket. This process involves encrypting their prediction (the embedding) so it cannot be read until a specific time has passed, preventing others from copying their signal before it's evaluated.
- Validation Process: Validators fetch these encrypted files after a delay (currently 300 blocks), decrypt them, and validate the contents.
- Ensemble Model: The system first trains a logistic regression model—a statistical method for predicting a binary outcome (e.g., price up or down)—on each individual miner's outputs. The outputs from these models are then fed into a final XGBoost model, a powerful machine learning algorithm that creates the final network-wide prediction.
- The codebase is remarkably small (around 1,400 lines of code), reflecting a minimalist design philosophy influenced by Bittensor core contributor Const.
The Incentive Mechanism: Rewarding True Contribution
- The most critical innovation in Mantis is its incentive mechanism, which directly ties rewards to a miner's unique contribution to the ensemble model's performance.
- Performance-Based Scoring: A miner's reward is calculated by measuring the performance degradation of the main XGBoost model when that specific miner's signal is randomized or removed.
- Example: If the network model has 55% accuracy and removing a specific miner's signal causes it to drop to 54.5%, that miner is responsible for a significant portion of the model's performance and receives a proportional share of the emissions.
- Barbarian states, "It's directly tied to the contribution to the network model since that's ultimately what we're working on optimizing."
- Strategic Implication: This design makes Sybil attacks economically unviable. If two miners submit the same signal, removing one has no impact on the network's performance, meaning the duplicate signal receives no reward. It incentivizes miners to develop unique, proprietary models.
The "Sub-Subnet": A Validator Optimization Testbed
- Barbarian introduces a novel concept he calls a "sub-subnet," which is essentially a controlled simulation environment designed to continuously optimize the Mantis validation mechanism itself.
- The Goal: This internal testnet uses synthetic, Gaussian noise-based data to simulate market conditions. This allows the team to test how the validator allocates rewards and resists adversarial strategies without risking real capital.
- The Process: Miners are incentivized to submit code that generates embeddings designed to degrade the network model's performance. This helps identify and patch vulnerabilities in the reward function.
- Actionable Insight: This represents a sophisticated approach to meta-optimization. By treating the validation mechanism as a machine learning problem, Mantis aims to build a system that is constantly learning and hardening itself against manipulation.
Future Vision: On-Chain Verification and Signal Monetization
- The long-term vision for Mantis is to create a verifiable, on-chain marketplace for its high-quality predictive signals.
- Trustless Verification: Mantis plans to move from a public randomness server to a private one. This will allow them to decrypt miner embeddings in real-time for inference, generate a prediction, and write the encrypted result to the Solana blockchain before the event occurs. This creates a publicly verifiable, tamper-proof record of the model's performance.
- Monetization Strategy: Once a consistent track record of high-quality signals is established, the plan is to auction them off. An on-chain auction would allow the highest bidder to receive the network's collective prediction for the next hour.
- Investor Takeaway: This creates a clear path to revenue generation. The value of the Mantis network (SN123) will be directly tied to the verifiable profitability of its signals, moving beyond speculative token value to a model based on tangible utility.
Community Q&A: Numerai, XGBoost, and Miner Quality
- Comparison to Numerai: Barbarian differentiates Mantis by highlighting its dynamic reward function. Unlike Numerai, where rewards can be less direct, Mantis is designed so that "any signal which improves the network model will be rewarded relatively swiftly," even if the edge is small.
- Why XGBoost? XGBoost was chosen because it outperformed other models like multi-layer perceptrons (MLPs) and linear regressions in initial trials. However, Barbarian acknowledges that better models likely exist and the "sub-subnet" is designed to help discover them.
- Miner Quality: Currently, only about 20-30 of the 256 registered UIDs are providing legitimate, high-quality signals. This highlights the ongoing challenge of separating signal from noise, which is the central focus of the Mantis incentive system.
Conclusion
Mantis's core innovation is its incentive mechanism, which is meticulously designed to isolate and reward genuine alpha in a noisy, decentralized environment. For investors and researchers, the project's success hinges on its ability to translate this theoretical advantage into a verifiable, on-chain track record of profitable signals.
