ORO

SN15 ORO runs a daily talent contest for robot personal shoppers, scores every performance, and keeps the footage as training material for a shopping assistant it plans to sell later.

The commodity, explained from zero

The contestants are shopping agents: small programs that, given a request and a budget, search a catalogue, compare prices and pick a product. Builders submit them. The subnet judges them on ShoppingBench, a benchmark the team built from 2.5 million real products, with new problems written every day so no agent can rehearse.

What the subnet produces today is the contest itself and its byproduct. Every run is recorded as a trajectory, a step-by-step log of what the agent looked at and decided. The front page counts 3,933,502 trajectories. The whitepaper says those logs trained a small model (4 billion parameters) from 18.0 to 42.7 percent on the benchmark, at "about 1/40th the cost of a frontier model".

Nobody buys anything yet. The roadmap puts "an agentic shopping assistant" under "Later". A waitlist endpoint exists in the API list. The materials name no customers and no price.

Why it is on Bittensor at all

The site's argument is openness: "AI agents are built and evaluated behind closed doors, so there is no way to compare them. ORO is an open arena." Independent validators run every agent, and their scores must agree. The comparison breaks here: a talent contest has one panel of judges, while ORO's judges are strangers who put up money for the right to score, and the prize is paid in the subnet's token rather than by a sponsor. Whether an open contest makes a better assistant than a closed lab, the materials do not show yet.

How the work gets done

Miners (the builders) submit one Python file with a single function, agent_main, and pay for their own model calls through Chutes or OpenRouter. Validators (the judges) run each file inside an isolated container against the day's problems, then hold a race on a hidden problem set for agents that qualified above a 55 percent threshold. An automated judge reads each agent's reasoning and scales its score by a coefficient from 0.3 to 1.0, so hardcoded answers earn less. The top agent, ranked by a difficulty-adjusted average of its last three races, takes "the large majority" of the emission (the subnet's share of newly minted TAO), a small protected share goes to the survivors, and the bottom 65 percent of each race are eliminated.

How you would know it works

The whitepaper page carries the numbers: base model 18.0 percent, ORO's trained model 42.7, the top miner's agent 77.3, a frontier model 64.0, with the models and 18,043 raw traces on Hugging Face. The front page's trajectory counter updates daily.

What is missing

There is no product a buyer can use and no price list. The chain contact team@oroagents.com appears nowhere on the site; the privacy policy gives support@oroagents.com instead. There is no status page and no support promise, only a blog post about a June 2026 attack. The miner docs do not state the registration cost (0.49 TAO on chain today) or give hardware requirements. Two version schemes coexist: the changelog reads v0.15.2 while GitHub releases stop at v1.0.2 from March.

Go deeper

• The front door, with counters
• Whitepaper with scores and models
• Races, elimination and the judge

Metaphor: a daily talent contest for robot personal shoppers. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.