Liquidity-Bounded Lending: Oracle-Fed Exposure Caps, Correlation-Aware Health, and Rate-Limited Liquidation for On-Chain Credit Pools

On-chain lending pools set collateral risk parameters per asset and per account, as though liquidations were independent events. A permutation test on Aave liquidations measured that this assumption understates the probability of simultaneous liquidations by about 301×. We describe a lending pool designed against that failure and report what a backtest does and does not support. The pool combines (i) a correlation-aware health function that scales loan-to-value and liquidation thresholds by the measured tail dependence of an account's collateral, with an asymmetry between borrow-side and liquidation-side tightening and a guard that reverts deposits which would worsen a debtor's health; (ii) a liquidity-bounded exposure cap, in which a separate oracle publishes each asset's trailing trading volume, the pool refuses deposits that push USD exposure (deposits plus both reserves) above κ times that volume, volume increases are limited in size and frequency, and stale volume data blocks new deposits; and (iii) a graduated liquidation stack (margin call, per-account cooldown, a protocol-wide leaky-bucket cascade brake that shrinks the close factor and lengthens cooldowns, and a Dutch-auction bonus). A 4.8-year backtest on nine Aave collateral assets finds that the correlation multiplier, by itself, does not outperform simply lending less, while per-asset liquidity caps cut bad debt by 9–34× at matched total debt. We report this negative result alongside the positive one, separate the elements with substantial prior art (Maker's global liquidation limit, Aave's rate-limited cap stewards) from the combinations we believe are new, and describe the mechanism precisely enough to serve as a technical disclosure.