Promptheus: AI-Driven Capital Allocation

Promptheus is an autonomous funding allocation system that combines token-weighted stakeholder signals with an LLM-based agent to create an autonomous capital distribution mechanism.

Core Mechanisms

(1) Influence Staking

  • Anyone can purchase and stake ALLO to influence the AI agent’s priorities with veALLO (vote-escrowed ALLO), where longer lock-up periods grant greater amount of non-transferable veALLO.
  • veALLO holders can then send text prompts to Promptheus, indicating mechanisms, builders, data, or opinions to consider.

(2) Autonomous Allocation Rounds

Every 7 days, Promptheus executes an allocation round:

  1. Pulls latest project updates, verified peer reviews, and onchain project metrics (if any)
  2. Combines this data with the guidance of staker prompts in order to numerically score all projects
  3. Posts transaction to distributes funds from pool proportional to scores
  4. Challenge period (3 days) where veALLO holders can veto Promptheus’s distribution. If veto threshold is not reached (e.g. 15%), the distribution is executed.

(3) Reputation & Revenue Sharing

  • Each allocation round updates staker reputation scores based on alignment of submitted prompts with the project scores for that round.
  • Reputation scores determine share of the future fees generated by project that are shared with the system (e.g. 0.5% of capital flows that are split with prompters on top of Allo protocol’s 0.5% fee)
  • This incentivizes ALLO stakers to provide high quality prompts to Promptheus.

Potential Enhancements

  1. Self-Improving Intelligence - Train Promptheus on successful allocation outcomes (fees generated) to improve its evaluation capabilities over time.
  2. Quadratic Weighting - Implement quadratic weighting for veALLO influence to prevent excessive concentration of power among whale token holders.
  3. Slashing - Implement a mechanism to slash veALLO stakes when prompts are flagged as malicious or consistently lead to poor allocation outcomes.

References

Partial inspiration from my previous work at Agentcoin:

3 Likes