Question 1

How is the quality of news analysis verified if each article is only checked by one validator?

Accepted Answer

Validators do not validate every article individually. Instead, each validator spot-checks a sample from every miner batch by re-running the same analyzer pipeline on those few articles and comparing the output against what the miner sent back. If the sample fails any of four validation tiers—enum match, deterministic hash match, embedding verification, or weighted composite—the entire batch is discarded. This audit-by-sample approach is cheap for validators but keeps miners honest because they know a few items will always be re-checked. Batches are accepted only if the sample fully reproduces; failing a single tier fails the whole batch.

Question 2

What happens if a miner times out or returns an incomplete batch?

Accepted Answer

Each miner holds a per-item lease from the validator. If the miner does not return a batch within the lease TTL, it goes back into the queue and is dispatched to another miner. Incomplete batches (fewer items returned than sent) trigger a size-mismatch penalty and the batch is rejected in full. The miner's batch size is then shrunk adaptively, and it enters a cooldown. Validators maintain per-miner adaptive windows that grow with demonstrated throughput and shrink after failures, so a consistently slow or incomplete miner gradually receives smaller batches until it either improves or times out entirely.

Question 3

Can a miner earn more by running multiple UIDs (unique identities)?

Accepted Answer

No. Each UID's share of emission is computed only from its own verified work. Running ten UIDs that each do one UID's worth of work earns one UID's pay total, split ten ways, and costs ten registration fees to do it. There is no multiplier for running more UIDs. The per-miner lease and dispatch window model also makes it impractical: a miner with one fast box gets more work faster than one split across ten UIDs, so consolidation is the dominant strategy.

Question 4

How does the miner know which articles to analyze or whether analysis is correct before sending it back?

Accepted Answer

The miner does not. It receives a batch from the validator with no ground truth attached, analyzes every item in the batch using its local copy of the analyzer pipeline, and returns the full batch to the validator asynchronously via a background thread. The validator then samples a few items, re-runs the exact same analyzer on them, and checks whether the miner's output reproduces. The miner has no feedback until the batch is accepted or rejected. This is what makes the audit work: the miner cannot game the sample because it does not know which items will be checked.

Question 5

What data does the miner need to download on first run?

Accepted Answer

The miner downloads approximately 44 GB of models on first run, including spaCy transformers, entity linking models (ReFinED), NER fusion engines (GLiNER, Flair), and embedding models (sentence-transformers). These must be version-aligned between miners and validators because validators re-run the same pipeline to verify determinism. The models are large because the pipeline runs four independent NER engines and fuses their outputs to achieve high-precision entity resolution across English and multilingual text.

Question 6

Is there a way to run a miner without an LLM API key?

Accepted Answer

No. Two of the four pipeline stages are LLM calls (extraction and reasoning), so an OpenRouter API key is required. The miner pays for its own inference, which is the main running cost after hardware. The cost scales directly with article throughput, so it encourages miners to optimize for both speed and accuracy. Validators also need their own API key and incur inference costs during spot-checks.

Question 7

What does 'verified volume' mean, and how is it converted to payment?

Accepted Answer

Verified volume is the count of items—articles, tweets, or messages—that a miner analyzed in batches that passed their validator audit. Each verified item earns the miner one point (or a weight based on content length in the article path). Those points are priced in USD at a fixed rate set by the subnet's emission schedule. The miner's total points for an epoch are converted to a percentage of available miner emission. If the network claims less than 100% of miner emission in total, each miner is paid exactly what its points are worth at the USD price, and the unclaimed remainder burns. If the network claims more than 100%, all shares scale proportionally and all emission goes to miners. Reputation acts as a multiplier: below a reputation threshold, the payment multiplier collapses toward zero; above it, it approaches 1.0.

AlphaRidge.ai

AlphaRidge.ai

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup

Participate

Participate

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup