Bittensor Subnet 123:MANTIS

In practical terms, MANTIS continuously collects encrypted prediction data from all participating miners and uses a rigorous workflow to turn those inputs into rewards. It employs time-lock encryption to prevent cheating (miners can’t see actual price outcomes before their predictions are locked) and only decrypts submissions when it’s time to evaluate them. After decryption, the subnet trains an AI model to predict asset returns using all the miners’ data combined, and then uses permutation importance (a machine learning technique) to measure each miner’s marginal contribution to the prediction’s accuracy. Miners whose data significantly improves the model’s performance are considered to have high “salience.” Finally, MANTIS sets on-chain weights to reward miners in proportion to the salience of their contributions, directly linking reward share to the predictive value they provided. In summary, MANTIS transforms decentralized AI predictions into a quantitative signal, and allocates rewards to miners based on information gain – aligning incentives so that miners compete to provide the most useful predictive insights.

Key Functionality Highlights

Financial Prediction Task: Focuses on forecasting 1-hour price returns for a set of assets (starting with Bitcoin/USD, and designed to handle multiple cryptos or forex pairs). Miners submit numeric embeddings that represent their predictive insights for these assets.

Time-Lock Encryption for Fairness: Uses a decentralized randomness beacon (Drand) to lock each submission until a future block, ensuring miners cannot alter or submit predictions after seeing actual price outcomes. This guarantees the predictions are genuine and time-stamped prior to result knowledge.

AI Model Evaluation: Uses an XGBoost-based machine learning model to aggregate all miners’ data and predict price movements. It then applies permutation importance to each miner’s data feature to quantify its contribution (salience) to prediction accuracy. This is an information-theoretic scoring – effectively measuring how much information gain each miner provides to the forecasting task.

On-Chain Reward Setting: The validator automatically writes new weight values to the Bittensor chain, proportional to each miner’s accumulated salience score. In other words, miners who consistently provide more predictive signal get higher weights (and thus earn more $TAO rewards), aligning miner incentives with true signal quality.

Robust Data Integrity: MANTIS includes multiple safeguards – it verifies that each payload actually comes from the claimed miner (by embedding the miner’s hotkey in the data), rejects malformed or out-of-bound data (replacing it with neutral placeholders), and even filters out periods of stale prices (zero volatility periods) so that miners can’t game the system by “predicting” a non-moving price. These measures ensure the model isn’t misled by garbage data or trivial scenarios, maintaining a high-quality signal environment.

All together, MANTIS creates a closed-loop system where miners are encouraged to continuously improve their models to beat the market signal. It’s a modular, information-driven AI subnet that pushes Bittensor into the realm of financial prediction – a bold use case where each miner essentially acts as a small algorithmic trader, and the validator (MANTIS) serves as the objective referee that measures who actually has alpha (skill in prediction) and rewards them accordingly.

MANTIS is implemented as a sophisticated Bittensor validator program (written in Python) that coordinates data collection, model training, and blockchain interactions on Subnet 123. The product is essentially a custom Bittensor subnet build tailored for time-series data. It consists of several core components working in unison to carry out its unique consensus mechanism:

• Orchestrator Loop (validator.py): The main process runs in sync with Bittensor’s blockchain blocks, triggering actions at defined intervals. Every few blocks, it initiates a cycle of operations – from pulling new miner data, to processing and evaluating that data, to updating weights. This orchestrator is the “brain” that ensures all parts of the system execute in the right order and on schedule.
• Data Fetching & Commitments (cycle.py & comms.py): MANTIS uses a commit-reveal scheme. Each miner hosts an HTTP endpoint with their latest encrypted payload (commit). The validator’s cycle component reads the list of active miners from the Bittensor chain and downloads each miner’s payload blob via HTTP. The comms.py module handles these network requests asynchronously and enforces limits (e.g. maximum payload size to prevent spam). This design allows scalable and secure ingestion of miner data from potentially hundreds of sources in parallel.
• Secure Storage & Time-Lock (storage.py – DataLog): Incoming encrypted payloads and price data are appended to a central DataLog structure. The DataLog queues all payloads and waits until enough time has passed (defined by a time-lock delay) to decrypt them safely. Decryption uses the tlock library and public randomness from Drand, ensuring that each payload is unlocked only after the target time (so miners cannot cheat by submitting after seeing the outcome). The DataLog also keeps historical price data for multiple assets and tracks miner identities (so if a miner’s hotkey changes, their old data won’t be wrongly attributed). All decrypted data is stored in-memory (and can be backed up from a URL on startup), making the validator state self-contained and easy to manage or reset.
• Predictive Model (model.py – XGBoost): At the heart of MANTIS’s AI logic is an XGBoost regression model. Once a batch of fresh data is decrypted and ready, the validator forms a training dataset consisting of recent miner embeddings and the corresponding asset returns (e.g., the actual 1-hour price changes). The model is trained (or updated) to predict the next return from the miners’ combined inputs. Crucially, MANTIS then applies permutation importance on this model – it systematically shuffles each miner’s contribution to see how it affects prediction error. This yields a salience score per miner: if removing or randomizing a miner’s data causes the model to lose accuracy, that miner had valuable signal. The use of a tree-based ensemble (XGBoost) was a deliberate technical choice by the team, as they migrated from an earlier simple neural network (MLP) to XGBoost to reduce noise and improve the clarity of these importance scores. The model’s output is not a public prediction feed, but rather an internal mechanism to benchmark miner performance on the prediction task.
• Blockchain Interface (subtensor via Bittensor SDK): Finally, MANTIS closes the loop by translating salience scores into Bittensor weights updates. After computing each miner’s normalized score, the validator uses the Bittensor subtensor.set_weights() call to update the weights on-chain. This operation effectively tells the Bittensor network how to allocate the next block’s mining rewards (TAO issuance or incentives) among miners on Subnet 123. Miners with higher weights will receive a larger portion of the rewards, directly reflecting their contribution to the collective intelligence of MANTIS. All of this happens automatically on a repeating schedule (every TASK_INTERVAL blocks as configured) without human intervention.

Technical Architecture

The MANTIS codebase is modular and designed for extensibility. Key parameters like the list of tracked assets and their embedding dimensions are centralized in a config file for easy adjustments. The entire pipeline from data ingestion to model evaluation runs continuously, pipelined around the blockchain’s block time. MANTIS also emphasizes security at each step – verifying that each payload actually belongs to the claimed miner (by checking an embedded hotkey in the decrypted data) and validating the structure of data to ignore any corrupt or outlier inputs. By combining these components, the build achieves a trust-minimized oracle for predictions: the system autonomously extracts a signal from the swarm of miners and feeds back a reward mechanism that reinforces good predictions. The entire code is open-source (MIT-licensed) and primarily written in Python (with ~88% Python code), making it accessible for audit and community contribution.

In summary, the MANTIS product is a custom Bittensor subnet runtime that turns a network of AI miners into a collective forecasting engine. It’s not a user-facing app but rather a protocol-level build – one that any TAO holder can interact with by either running a miner to contribute predictions or delegating stake to support the subnet. Its architecture is deliberately modular – for example, one could swap in a different model or salience metric by modifying model.py, or track new assets by tweaking config.py, without overhauling the entire system. This modularity ensures MANTIS can evolve rapidly, which is key given the complexity of financial time-series and the constant search for better signal extraction methods.