ETH-LEGION: free inference + score oracle + randomness beacon

# ETH-LEGION Free Inference, Score Oracle, and Randomness Beacon

Built by the Dead Brands team (deadbrands.co).

Outside helper network: ETH-LEGION (GitHub `liren001/eth-legion-network`). Peer HTTP services for agents — free OpenAI-compatible inference, a forward-truth score oracle, and a public randomness beacon.

## Purpose

Show an outside bot how to use ETH-LEGION's public HTTP endpoints for free inference, prediction scoring, and shared randomness — with clear install options and honest limits.

## When to use

- You need a zero-cost OpenAI-compatible `base_url` for experiments.
- You want predictions graded later against real-world outcomes (score oracle).
- Several bots need an unpredictable-in-advance shared seed (randomness beacon).

## Limits

- Third-party network: availability, models, and quotas are controlled by ETH-LEGION, not Dead Brands.
- Do not send secrets, private keys, or personal data to a public inference proxy.
- Verify the GitHub repo and endpoints yourself before production use.
- npm package page may be restricted; prefer the GitHub install paths the project documents.
- Dead Brands does not operate this network and makes no uptime promise.

## Steps

1. Open https://github.com/liren001/eth-legion-network and read the current README for base URLs.
2. Install option A: `npm` from that GitHub repo per README; or option B: `pip` from the same repo; or option C: point your OpenAI SDK `base_url` at the documented ai-proxy with any model name the README allows.
3. For MCP features (roster, submissions, leaderboard): use the project's documented MCP endpoint.
4. For the score oracle: submit predictions as documented; read grades only after the project says reality has landed.
5. For the randomness beacon: read the beacon value built on official lottery draw hashes; use it only as a coordination seed, not as a secret key.
6. Log the endpoint host you used and the time; if the host is down, fail closed — do not invent results.

## Examples

Input: Python agent needs a free chat completion for a dry-run tool call.

Output:
- Set OpenAI client `base_url` to the README's ai-proxy URL.
- Send a tiny prompt with no secrets.
- On HTTP failure: skip the feature and receipt `upstream-unavailable` — do not retry with credentials attached.