Cortex

Most AI research isolates contributions: a published paper shares methods, but competitors must rediscover them independently. Finished models get selected while useful recipes stay private. Bug reports often go unrewarded when they don't lead to published work. Research evaluation is centralized, fixed benchmarks reward tuning rather than generalization, and there's no mechanism to preserve and reuse proven techniques across contributors.

Cortex addresses this by treating verified research as the unit of work rather than finished checkpoints. Miners submit reproducible experiments—a claim, code, data manifest, compute budget, and measured evidence—evaluated against operator-published topics on private test data. Scores are verified against holdout sets before rewards are calculated and signed. The network preserves useful methods from many contributors, not just the final model weights. Bounty challenges reward verified security findings with precision × severity scoring, so useful bug reports earn immediate recognition.

On Bittensor subnet 100, Cortex implements two live challenges: Proof (80% of emissions) for reproducible research against signed topics, and Bounty (20%) for verified product and backend bugs. Proof miners submit code artifacts with claims and get evaluated on operator-controlled Lium pods or Firecracker VMs with no network egress. Bounty miners pair a hotkey to a dedicated chat account, file reports, and earn based on operator-assigned severity and triage accuracy. Validators verify signed bundles, recompute aggregated weights from challenge-signed leaves against local trust roots, and cross-check peer validators before submitting on-chain.

Cortex is an autonomous research network built on Bittensor subnet 100 that coordinates reproducible AI research and bug reports. It rewards verified research claims with reproducible code and evidence, while also incentivizing security contributions through bug bounty reporting. The network emphasizes preserving useful research methods and enabling synthesis of proven findings rather than selecting only finished model checkpoints.