AlphaRidge.ai

# Alpharidge AI (Subnet 45)

Subnet 45 reads the world's financial news and turns it into structured market data. About 290,000 articles a day, from 12,997 sources. It's the data layer behind **[AlphaRidge](https://alpharidge.ai)**, which is where that data goes to be used.

Almost all of that work is done by miners. A miner reads an article and produces a full analysis of it: sentiment per asset, entities resolved to Wikidata IDs and tradeable tickers, economic data points, contagion chains, narrative tags and embeddings.

Validators hand out the work in batches and spot-check a sample from each one, re-running the pipeline themselves on those few articles without seeing what the miner sent back. If the sample doesn't reproduce, the whole batch is thrown out and earns nothing. That keeps the volume honest without paying to do it twice.

| | |
|---|---|
| Articles analyzed to date | **16.9 M** |
| Analyzed per day | **~290,000** |
| Distinct source domains | **12,997** |
| Real-world events clustered | **2.6 M** |
| Tickers with linked coverage | **2,138** |
| Assets tracked by the extractor | **2,314** |

<sub>Measured 2026-08-29. Live throughput, the analysis stream and the miner leaderboard run continuously on the subnet observatory at [alpharidge.ai](https://alpharidge.ai).</sub>

**New miner?** Go straight to the [Quickstart](#quickstart-run-a-miner).

---

## Table of contents

- [What this powers](#what-this-powers)
- [Quickstart: run a miner](#quickstart-run-a-miner)
- [How you earn](#how-you-earn)
- [Where the rest of the emission goes](#where-the-rest-of-the-emission-goes)
- [What a miner actually does](#what-a-miner-actually-does)
- [How validation works](#how-validation-works)
- [Dispatch](#dispatch)
- [Configuration](#configuration)
- [Running a validator](#running-a-validator)
- [Architecture](#architecture)

---

## What this powers

SN45 does two things: it reads every article, and it works out how those articles relate to each other. AlphaRidge is the product built on both.

| Layer | What it does | Where |
|---|---|---|
| **Perception** | Read and analyze every article; verify every analysis | **SN45** |
| **Cross-article intelligence** | Cluster articles into real-world events; match events to long-running market narratives | **SN45** |
| **Product** | Chart, Screener, Watchlists, Alerts, Backtesting, AI assistant, public API | AlphaRidge |

In the product, miner output shows up as:

- **News marks on the price chart**: every analyzed article plotted against the tape, ranked by impact.
- **Screener columns**: 7-day sentiment, article count, max impact, bull/bear split and 30-day event count, screenable across the whole equity universe.
- **A per-symbol news feed** built on real ticker linkage from entity resolution, not a text match on the company name.
- **An events browser**: articles grouped into the real-world event behind them, so one story is one row instead of forty.
- **A public sentiment and events API**, and an AI assistant that can read the same feed.
- **A live subnet observatory**: throughput, the analysis stream, trending events and a miner leaderboard.

---

## Quickstart: run a miner

### 1. Hardware

Miner and validator run the same analyzer, so they need the same box.

| | Minimum | Comfortable |
|---|---|---|
| GPU | 8 GB VRAM | 24 GB+ |
| RAM | 16 GB | 32 GB |
| Disk | 60 GB free | 100 GB |
| CPU | 8 cores | 16+ cores |

About 44 GB of models download on first run. The NER stage is CPU- and Python-heavy, not purely GPU-bound. Core count is what usually sets your throughput, and throughput is what you are paid for.

### 2. Install

```bash
git clone https://github.com/Team-Rizzo/alpharidge-ai.git
cd alpharidge-ai
python3.12 -m venv .venv && source .venv/bin/activate
./install.sh          # CUDA 12.8 default
```

For a different CUDA build: `TORCH_INDEX=https://download.pytorch.org/whl/cuXXX ./install.sh`

<details>
<summary>Manual install (equivalent, if you would rather not run the script)</summary>

```bash
python -m pip install --upgrade pip setuptools wheel

# 1. PyTorch: match the CUDA build to your driver (see https://pytorch.org):
pip install "torch>=2" --index-url https://download.pytorch.org/whl/cu128

# 2. The rest of the stack (the spaCy en_core_web_trf model is pinned in requirements.txt):
pip install -r requirements.txt
pip install -e .

# 3. ReFinED (Amazon entity linker) is not on PyPI. Install it with --no-deps so it
#    does not downgrade torch/transformers, then add its small runtime deps:
pip install --no-deps "git+https://github.com/amazon-science/ReFinED.git@V1"
pip install ujson nltk Unidecode lmdb prettyprint
```
</details>

### 3. Configure

```bash
cp miner_env.example .miner_env
```

**For a standard setup you only need to set one value: `API_KEY`.** Everything else in the template is pre-filled.

`API_KEY` is your own [OpenRouter](https://openrouter.ai/keys) key (`sk-or-...`). Two of the four pipeline stages are LLM calls, so you pay for your own inference. It's your main running cost after hardware and it scales directly with how many articles you get through, so budget for it.

### 4. Register and run

```bash
btcli subnet register --netuid 45 --wallet.name <coldkey> --wallet.hotkey <hotkey>

.venv/bin/python -m neurons.miner \
  --netuid 45 \
  --wallet.name <coldkey> \
  --wallet.hotkey <hotkey> \
  --logging.info
```

Optional: `--axon.external_ip` and `--axon.external_port` if you are behind NAT or a proxy.

> **If your axon sits behind nginx**, set `underscores_in_headers on;`. Bittensor sends `bt_header_*` request headers, and nginx silently strips underscored headers by default. Your miner will look healthy and never receive a single batch.

### 5. Confirm it is working

First run downloads ~44 GB of models, so give it time before you judge anything.

- Your logs should show batches arriving, then an analysis time of roughly 12–18 seconds per article.
- Batches arrive on a lease. If you do not return one in time i