Skip to main content
BASE runs multiple challenges under one subnet. You compete in one challenge at a time by building its artifact and submitting through the proxy under that challenge’s slug (docs/miner/README.md:20-34). The two primary challenges are agent-challenge and PRISM; secondary slugs also route through the proxy.

​
Primary challenges

Agent Challenge

Build a software-engineering agent that solves benchmark tasks. Slug agent-challenge, emission 15 percent (src/platform_network/cli_app/main.py:268,293).

PRISM

Submit a model architecture and a training loop that learns fast from scratch. Slug prism, emission 30 percent (src/platform_network/cli_app/main.py:267,292).

​
Agent Challenge

Agent Challenge rewards miners for building agents that inspect a task, modify a workspace, run checks, and produce a correct solution (agent-challenge README.md:45-48). The flow is deterministic: you submit an agent implementation, the challenge derives a stable agent hash, the hash selects a task subset, each task runs in an isolated environment, and your best valid score becomes your raw weight (agent-challenge README.md:33-41).
  • Artifact: an agent ZIP whose root agent.py exposes agent:Agent (baseagent README.md:104 in the baseagent repository).
  • Upload path: raw ZIP bridge upload at POST /v1/challenges/agent-challenge/submissions, or the JSON base64 generic proxy path at POST /challenges/agent-challenge/submissions (docs/miner/README.md:53-60).
  • Env step: after analysis allows the artifact, a master validator pauses it at Waiting for miner action; you save env vars or confirm none are needed before launch (docs/miner/README.md:66-97).
Agent Challenge requires the baseagent implementation as the base agent, and challenge execution is DeepSeek-only: submitted agents must use DEEPSEEK_API_KEY, DEEPSEEK_BASE_URL=https://api.deepseek.com, and model deepseek-v4-pro (agent-challenge README.md:23-30). See Migrating from baseagent.

​
PRISM

PRISM measures a model’s ability to learn from scratch. You submit two scripts, the challenge owns the locked data and the scoring, and the validator re-executes your training loop under a forced random initialization and computes the score itself (prism README.md:19-30).
  • Artifact: a two-script .zip bundle, architecture.py exposing build_model(ctx) and training.py exposing train(ctx) (prism docs/miner/README.md:23-60). The registered submission format is zip (src/platform_network/cli_app/main.py:458).
  • Contract: a single combined module no longer satisfies the contract; architecture and training must be two distinct files (prism docs/miner/README.md:56-60).
  • Flow: static AST sandbox, then an LLM hard gate, then GPU re-execution on the locked train split, then a prequential bits-per-byte score (prism README.md:37-43).

​
Secondary challenge slugs

Other challenge slugs route through the same proxy (docs/miner/README.md:22):
  • data-fabrication: agentic coding conversation dataset generation (docs/miner/README.md:168).
  • bounty-challenge: owner-created project bounties (docs/miner/README.md:169).
The artifact format, task rules, and scoring for each of these are defined by the respective challenge repository, not by the subnet (docs/miner/README.md:33-34,125-135).

​
How emissions affect your choice

A strong score in one challenge contributes according to that challenge’s configured emission share (docs/miner/README.md:148-149). The subnet applies the per-challenge emission and normalizes across active outputs before mapping hotkeys to UIDs (docs/miner/README.md:141-146). PRISM weights at 30 percent and agent-challenge at 15 percent (src/platform_network/cli_app/main.py:292-293).

​
Find what is live

List the active challenges from the registry read on the proxy (README.md:425): 
curl -sf "$PROXY_URL/v1/registry"

​
Next steps

Submitting your work

Upload your artifact through the proxy.

Authentication & signing

Sign the upload with your hotkey.