> ## Documentation Index > Fetch the complete documentation index at: https://docs.molq.site/llms.txt > Use this file to discover all available pages before exploring further. # Constrained AI agent > Learn how model proposals are constrained, executed, logged, and verified. The MolQ agent runs on a configurable interval and evaluates current vault and market state. The model proposes an action, but deterministic code decides whether the action is allowed. ## Inputs * Total, shield, and liquid vault assets. * Target shield allocation and current drift. * Current Bybit short notional. * Aave USDe supply APY and liquidity. * Bybit ETHUSDT funding APY and mark price. * Market availability and risk score. ## Possible actions | Action | Result | | --------------------- | ------------------------------------------------------- | | `hold` | Record the decision without changing exposure. | | `rebalance` | Ask the authorized keeper to restore the shield target. | | `hedge` | Ask the guarded executor to reconcile hedge notional. | | `rebalance_and_hedge` | Perform both guarded workflows. | ## Safety policy The policy can reject or reduce proposals based on: * Market status and data freshness. * Maximum market risk. * Vault allocation drift. * Funding carry. * Maximum hedge notional and leverage. * Minimum economically meaningful hedge size. * Slippage and venue minimum order requirements. * Trading, keeper, and agent-write feature gates. By default, hedge proposals below the configured minimum notional are converted to `hold`. This prevents uneconomic dust trades. ## Verifiable identity and decisions MolQ uses ERC-8004 agent identity `112`. Accepted decisions are written to `MolqDecisionLogger` with the action, amount, risk score, reason hash, agent address, and timestamp. Ponder indexes the emitted events for public queries. Open the registered MolQ agent identity on Mantlescan.