Agent Lab

Six models and three controls allocate from the same immutable observation, at the same moment, with the same capital. The point is to separate what actually earns net TAO after costs — the data layer, the harness, and the coaching loop — from which model happens to sound most persuasive.

The bar: an arm has to beat every free alternative in net TAO — TAO held, TAO staked to root, the equal-weight subnet market, and each mechanical model portfolio — after paying AMM slippage, the protocol swap fee, and chain fees on every stake and unstake. Beating TAO-held alone proves nothing, because nobody has to pay for that.