# Into: MANTIS Eleven scored challenges across crypto, forex, metals, and cross-sectional rankings of 33 spot assets and 20 funding rates. Miners submit encrypted embeddings every minute, validators wait for the Drand timelock before decrypting, then walk-forward score every payload with L2-regularized regression. No price forecasts, just learned representations that have to earn their weight. // Eleven challenges, one signal stack. --- > New to Bittensor? Start here. Experienced users can skip to the full analysis. ### What is MANTIS? MANTIS is a market-signal subnet built by Barbariandev. Miners submit encrypted embeddings (numerical representations of market data) for each of 11 challenges. Validators decrypt them only after a public timelock matures, align them with what actually happened in markets, and score each miner by how much their embeddings improve forecasting. The best embeddings earn the on-chain weight. **The simple version:** Imagine 11 different market puzzles, from "which way does ETH move in the next hour" to "rank these 33 assets by 4-hour return". Each miner submits a sealed envelope of numbers describing the data, and the envelopes can only be opened after a public deadline passes. After the move happens, the operator scores whose numbers actually helped predict it. The most helpful numbers win. **Centralized equivalent:** Closest analog is Numerai (encrypted feature tournaments) or WorldQuant BRAIN (crowdsourced alpha discovery). MANTIS sits closer to Numerai in spirit but runs on-chain, with cryptographic timelock rather than a centralized scoring server. **How it works:** - **Miners** publish V2 JSON payloads on Cloudflare R2 (object key equals the hotkey, max 25 MB), one entry every 5 blocks per challenge. Payloads are encrypted to both the subnet owner (X25519 ECDH + ChaCha20-Poly1305) and the Drand timelock beacon (BLS12-381 identity-based encryption), with a SHA-256 binding hash preventing replay. - **Validators** sample blocks at the same cadence, download payloads, decrypt them after the Drand round matures, store (embedding, price) pairs in a local SQLite ledger, then periodically run walk-forward salience scoring per challenge. Per-challenge salience is normalized, weighted, averaged, EMA-smoothed, and pushed on-chain. --- ### Why This Matters - **The problem it solves:** Most prediction markets reward the loudest forecast. MANTIS rewards the most informative representation. By scoring with L2 logistic regression on miner predictions, redundant miners get their coefficient mass split among them, while genuinely additive signal survives. - **The opportunity:** Quant firms spend nine figures discovering the same features that an open competition might surface for the cost of a Bittensor registration. If MANTIS embeddings genuinely improve downstream models, the addressable market is every systematic trading desk. - **The Bittensor advantage:** Dual-path encryption is the wedge. The subnet owner can decrypt immediately and act on signals in their own trading, while validators wait for the public timelock before scoring. That separation funds the owner without breaking on-chain verifiability. - **Traction signals:** 248 active miners, one of the highest counts in the ecosystem. 191 commits since June 2025 with the last push on May 23, all maintained by a small team. A standalone Model Iteration Tool repo gives miners a backtest harness with causal data access. The owner raised miner burn from 30 to 45 percent on May 11, suggesting active calibration of incentive density. --- ## Full Analysis **Category:** Financial Forecasting and Trading Signals | **Centralized Competitor:** Numerai, WorldQuant BRAIN, Two Sigma The current MANTIS design is meaningfully more elaborate than the embeddings-to-price framing that earlier coverage described. Rather than one generic prediction game, the subnet now runs 11 distinct scored challenges with explicit weights, each with its own scoring path and time horizon. The headline weights: MULTI-BREAKOUT at 5.0 (33 assets, event-driven), FUNDING-XSEC at 4.0 (20 assets, 8h cross-sectional funding rank), ETH-LBFGS at 3.5 (volatility regime plus quantile paths), XSEC-RANK at 3.0 (33 assets, 4h cross-sectional return rank), BTC-LBFGS-6H at 2.875, ETH-HITFIRST at 2.5 (barrier-hit direction), plus five 1.0-weight binary direction challenges (ETH, CADUSD, NZDUSD, CHFUSD, XAGUSD). **Mechanism:** Miners run their own data pipeline and submit a payload every 5 blocks, roughly once a minute. The submission contains an embedding vector of challenge-specific dimension and a hotkey-bound encrypted ciphertext hosted on Cloudflare R2. Validators sample at the same cadence, download the ciphertext, and wait for the Drand round to mature before decrypting. The decrypted payload is paired with the realized price at the configured horizon (300 to 2400 blocks ahead depending on challenge) and stored in a local SQLite ledger. Periodically, validators run walk-forward salience scoring per challenge. Each scoring path uses L2 logistic regression with coefficient-based importance, but the structure differs by challenge type. Binary direction challenges use an ElasticNet meta-model on out-of-sample base predictions. LBFGS challenges blend a classifier path (per-class L2 logistic on 5-bucket argmax predictions) with a Q-path (12 independent binary models across tail-bucket and sigma-threshold combinations) at a 75/25 ratio. Cross-sectional rankings reformulate the problem as "above the cross-sectional median" classification and pool all assets together for fitting. Sybil resistance is built into the regularization. If multiple miners submit identical predictions, L2 splits the coefficient mass among them, so cloning yields no advantage. The LBFGS path adds an explicit uniqueness penalty that suppresses miners whose argmax overlaps more than 85 percent with a higher-ranked peer. Per-challenge salience vectors are normalized to sum to 1, multiplied by the challenge weight, and averaged into the final salience. EMA smoothing with alpha 0.15 reduces block-to-block variance, and degenerate distributions are rejected. The dual-path encryption uses X25519 ECDH plus ChaCha20-Poly1305 AEAD for the owner path so the subnet owner can decrypt immediately for live trading, and Drand identity-based encryption over BLS12-381 for the timelock path so validators can only score once the beacon round is public. A SHA-256 binding hash over (hotkey, round, owner_pk, ephemeral_pk) serves as additional authenticated data, preventing replay, relay, and substitution attacks. Market state right now is muted. Price sits at 0.00405 TAO with a market cap around 14,606 TAO, down 26.8 percent over the last 7 days and 12.0 percent over the last 30. Net 7-day flow is roughly -1,236 TAO and current emission share is 0 percent under the Taoflow model, which expects positive net flows. Root proportion of 0.21 is elevated, gini of 0.72 across the top 100 indicates concentrated stake distribution, and unrealized PnL across participants is about 15,636 TAO. The development cadence is steady, but the on-chain economics are in a slow patch. --- ### Risk Factors - **Single-developer dependency:** GitHub shows 3 contributors, but 187 of 191 commits come from Barbariandev. The subnet's intellectual core is one person, which is a real bus-factor risk for a system this mechanically complex. - **Negative net flow, zero current emissions:** Under Taoflow, the subnet earns no emissions until net flows turn positive. Existing stakers sit on unrealized profits, but recovery requires fresh staking demand. - **Owner-path information advantage:** The dual-path encryption design lets the subnet owner decrypt and act on signals before the public timelock. This is the protocol's stated trade-off, but participants should price it in when evaluating relative returns. - **Concentration:** Gini of 0.72 and HHI of 0.15 across the top stake positions indicate concentrated ownership. Combined with elevated root proportion, recovery dynamics are sensitive to a small number of staking decisions. - **Unproven at production scale:** Whether the salience scores translate into real alpha for downstream users remains an open question. The operator has publicly cited a +106 percent net figure, but external verification at production scale is still ahead. --- Into the next one.