MANTIS

# MANTIS

Bittensor Subnet 123, The Ultimate Signal Machine

---

## Architecture

Validators sample miner payloads every `SAMPLE_EVERY` blocks, decrypt them after a timelock maturation window, and store (embedding, price) pairs in a SQLite database. Periodically, walk-forward scoring computes per-hotkey salience for each challenge, aggregates across challenges by weight, applies EMA smoothing, and sets on-chain weights.

```mermaid
graph TD
    subgraph validator["validator.py"]
        A["Sample block"] --> B["cycle.get_miner_payloads()"]
        B --> C["ledger.append_step()"]
        A --> D["Periodic weight calc"]
        D --> E["ledger.iter_challenge_training_data()"]
        E --> F["model.multi_salience()"]
        F --> G["EMA smooth + set_weights()"]
    end

    subgraph ledger["ledger.py (SQLite)"]
        H["challenge_data"]
        I["raw_payloads"]
        J["drand_cache"]
    end

    subgraph external["External"]
        K["Miners (R2)"]
        L["Drand beacon"]
        M["Subtensor"]
        N["price_service.py → R2"]
    end

    C --> H
    C --> I
    E -- reads --> H
    B -- downloads --> K
    I -- decrypt via --> L
    G -- writes --> M
    B -- reads commits --> M
    N -- publishes --> H
```

---

## Challenges

All challenges are defined in `config.py` under `CHALLENGES`. Each specifies a `ticker`, `dim`, `blocks_ahead` (forward horizon in blocks at 12s/block), `loss_func` (scoring dispatch key), and `weight` (relative importance in final aggregation).

| Challenge | Ticker | dim | Horizon | loss_func | Weight | Description |
|---|---|---|---|---|---|---|
| ETH-1H-BINARY | `ETH` | 2 | 300 (1h) | `binary` | 1.0 | Binary direction prediction |
| CADUSD-1H-BINARY | `CADUSD` | 2 | 300 | `binary` | 0 | " |
| NZDUSD-1H-BINARY | `NZDUSD` | 2 | 300 | `binary` | 0 | " |
| CHFUSD-1H-BINARY | `CHFUSD` | 2 | 300 | `binary` | 0 | " |
| XAGUSD-1H-BINARY | `XAGUSD` | 2 | 300 | `binary` | 0 | " |
| ETH-HITFIRST | `ETHHITFIRST` | 3 | 500 | `hitfirst` | 0 (sunsetting) | Barrier-hit direction |
| ETH-LBFGS | `ETHLBFGS` | 17 | 300 (1h) | `lbfgs` | 3.5 | Volatility regime + quantile paths |
| BTC-LBFGS-6H | `BTCLBFGS` | 17 | 1800 (6h) | `lbfgs` | 2.875 | " |
| MULTI-BREAKOUT | `MULTIBREAKOUT` | 2/asset | event | `range_breakout_multi` | 5.0 | Range breakout continuation/reversal (33 assets) |
| XSEC-RANK | `MULTIXSEC` | 1/asset | 1200 (4h) | `xsec_rank` | 3.0 | Cross-sectional return ranking (33 assets) |
| FUNDING-XSEC | `FUNDINGXSEC` | 1/asset | 2400 (8h) | `funding_xsec` | 4.0 | Cross-sectional funding rate ranking (20 assets) |
| FLOW-BTC | `FLOW` | 28 | horizon-driven, 1–336h | `flow` | 1.020 (5% since 2026-10-26; rising toward 25% as the field grows) | Capital-at-risk BTC bracket trades, evidence-gated payment |

TRADE-MIX is deprecated as of the FLOW launch (2026-08-17): off the
active roster, historical challenge data purged from validator
datalogs on open.

---

## Scoring

### Per-challenge scoring

Each `loss_func` has its own scoring path. All use L2 logistic regression; importance is coefficient-based except the LBFGS classifier path, which uses out-of-sample permutation importance.

**Binary** (`binary`) — Walk-forward with ElasticNet meta-model. Feature selection: per-miner L2 logistic on first half, AUC on second half, select top-$K$ (default 50). Meta-model: ElasticNet logistic (L1 ratio 0.5) on OOS base-model predictions across walk-forward segments. Importance = $|\beta_j|$. Segments weighted by recency.

**LBFGS** (`lbfgs`) — Miners whose bucket probabilities correlate above 0.85 (lag-tolerant, consensus-removed, trailing 20,000 rows) are merged into one median signal. Two independent scoring paths blended 75/25:
- *Classifier path* (`compute_linear_salience`): per-class L2 logistic regressions on 5-bucket argmax predictions, in end-anchored walk-forward segments. Importance = out-of-sample balanced-accuracy drop when a miner's column is permuted. Vectorized balanced accuracy evaluation.
- *Q-path* (`compute_q_path_salience`): 12 independent binary L2 logistic models (one per tail-bucket / sigma-threshold combination). Importance = averaged $|\beta_j|$ across sub-models.

Both paths are individually top-$K$ truncated and renormalized before blending.

**HITFIRST** (`hitfirst`) — Two L2 logistic regressions on logit-transformed miner probabilities: one for up-barrier-hit ($y=1$ if price hits $+\sigma$ first), one for down-barrier-hit. Importance = $|\beta_j^{\text{up}}| + |\beta_j^{\text{down}}|$. No walk-forward — single fit on all valid samples.

**MULTI-BREAKOUT** (`range_breakout_multi`) — Operates on completed breakout events (not time series). Two-stage: (1) Empirical AUC gate — per-miner AUC on $P_{\text{continuation}}$ vs realized label, requiring AUC > 0.5 and ≥ 2 temporal episodes. (2) L2 logistic on z-scored miner predictions with episode-balanced sample weighting (each temporal episode gets equal total weight regardless of event count). Importance = $|\beta_j|$.

**XSEC-RANK** (`xsec_rank`) — Cross-sectional binary reformulation: label = 1 if asset's forward return exceeds the cross-sectional median. All assets pooled ($N_{\text{assets}} \times$ sample multiplier). Walk-forward meta-model: feature selection by per-miner univariate AUC, top-$K$ (default 20) selected, L2 logistic meta-model. Importance per segment:

$$
w_j = |\beta_j| \cdot \max\!\Big(\frac{\text{AUC}_{\text{meta}} - 0.5}{0.5},\; 0\Big)
$$

Segments aggregated with exponential recency weighting.

**FUNDING-XSEC** (`funding_xsec`) — Same structure as XSEC-RANK but on funding rate changes instead of price returns. Embargo = $\max(\text{LAG}, \text{ahead})$ with explicit `train_cutoff = val_start - ahead` to prevent label leakage from forward-looking labels. Stale miners (temporal std < $10^{-4}$ per asset column) zeroed before pooling.

**FLOW** (`flow`) — Not a regression challenge: miners submit bracket trades (direction, Kelly fraction, stop, two targets, horizon) across four horizo