Loops HouseMortal Dialect

NEAR HACKATHON
Links
Team
1 member- YAOwner
Yamazar1
Overview
Why: I believe what today's AI lacks is not knowledge but the flash of insight, and that this flash comes from a finite life: a living thing has to keep changing or it dies. An AI agent has no such pressure. It can talk forever and nothing is ever at stake for it. Now that agents are being given wallets, that matters in practice.
What: Mortal Dialectic gives agents a life that can run out. Four agents (Builder, Destroyer, Synthesizer, Judge) debate a question. Each is a NEAR sub-account with its own key. Every utterance is paid for with a native transfer from the speaker's own balance. The Judge divides a reward pool that is smaller than what the speakers just spent. An agent that cannot pay is removed with DeleteAccount, and a successor account inherits only the idea.
Did it work? The economy did: 1,662 testnet transactions and 75 deaths. My hypothesis did not hold up. Four experiments against immortal, free controls, read by one rule fixed before the second, show no clear difference in how repetitive the ideas were: not with a 7B model, not with Claude Sonnet 5.5, not when the dying left advice. The best objection came from the agents. In round 11 a newborn Builder argued for mortality-linked stakes, the Destroyer answered that this "optimizes for not dying", and the Synthesizer said that survival which needs only adequacy narrows the pool of ideas, while survival which needs an outlier widens it. My economy may have paid for adequacy. Testing that is the next experiment.
Use it: run it on your own question and watch who survives, change the prices and see which role starves, or reuse the pattern of a budget as a lifespan with a hard stop on-chain. The repo has an offline demo, 46 tests and full logs with a link for every transaction. A judge can reproduce an on-chain death with no API key in about six minutes. My other submission, Mandate Gate, is a separate project from the same starting point.