Skip to content
Enmonetic

Enmonetic Protocol

DRAFT

One engine under every product.

Particle swarm optimisation splits liquidity across vetted lending markets so that our own deposits move rates as little as possible. Before any allocation, the Collateral Risk Analyzer decides which pools are allowed and how much each may receive. It runs Lemonaire® today and the liquid part of Earn and Yield-as-a-Service.

00 · Flow of funds

How capital moves through the engine.

Deposits enter an ERC-4626 vault and the depositor receives shares. The risk layer vetoes unsafe pools, the PSO allocator splits the rest across approved markets, and yield flows back into the vault — raising the value of every share.

ILLUSTRATIVE MODEL
How capital moves through the engine.A depositor sends USDC to an ERC-4626 vault and receives shares. The vault passes liquidity through the CRA risk gates to the PSO allocator, which splits it across approved markets; a pool that fails a gate gets zero. Yield flows back to the vault and raises the value of each share.USDCSHARESYIELD → SHARE VALUE ↑VETODEPOSITORWalletholds sharesERC-4626 VAULTVaultasset in · shares outCRA · RISK LAYER8 gatesscore → max allocationALLOCATORPSOoptimal splitMetaMorpho vault A38%Aave v3 USDC36%Morpho Blue market B18%Idle buffer8%Pool X · gate failed0%
Illustrative split. Pools that fail a hard gate receive zero allocation.

01 · The problem

More liquidity means a lower rate.

Supplying to a lending market lowers its rate. An allocator that simply chases the highest quoted rate pushes that rate down and leaves yield on the table. The engine spreads liquidity so that no single market is saturated and the blended rate stays as high as possible.

02 · How we got here

From exhaustive search to a swarm.

  1. Replaced

    Exhaustive search

    Simulations plus search over every split. Exact for two or three markets, exponentially expensive as markets are added.

  2. Rejected

    Gradient methods and Bayesian optimisation

    SLSQP, L-BFGS-B and Bayesian optimisation need gradients or a smooth model that real markets do not provide, and were not accurate enough.

  3. In use

    Particle swarm optimisation

    Needs no explicit APR(L) formula and no gradients, and scales to many markets.

03 · Try it

Watch a swarm balance yield against risk.

Set each market's APY, pool size and CRA risk score. The swarm looks for the highest blended yield while keeping every pool within the share its risk score allows. Synthetic parameters, computed in your browser.

04 · Continuous rebalancing

Re-solved daily and whenever the risk layer moves.

  • New deposits are allocated so that the added liquidity lowers blended APY as little as possible.
  • The base cycle is a daily rebalance, plus unscheduled rebalances triggered by CRA events.
  • A rebalance executes only when the expected gain exceeds gas and slippage.

05 · Risk layer

The Collateral Risk Analyzer.

Enmonetic supplies stablecoins; it does not borrow. The real risk sits in the collateral behind the loans our liquidity funds. Before every allocation CRA asks one question: what is the actual risk of losing principal in this pool, given the quality and concentration of that collateral?

18 Apr 2026 · rsETH on Aave · ~$292M

Why price and health factor are not enough

The signal was not in the health factor or the price. It was in the bridge design, the speed of collateral inflows and the addresses behind them.

  1. T₀

    Unbacked rsETH is minted through a bridge secured by 1-of-1 verification.

  2. T₀ +

    The attacker supplies it as collateral and borrows WETH against it.

  3. Positions look healthy: health factor 1.01–1.03, oracle price normal.

  4. T₀ + ~46 min

    The protocol is paused.

Four independent contours

  • 01

    Structural

    Concentration (HHI by market, asset and cluster of correlated assets), oracle quality, asset profile — bridge, redemption, proof of reserve — and market parameters.

    rsETH case · 1-of-1 bridge on a material collateral asset → hard gate.

  • 02

    Positional

    Distribution of debt across health-factor buckets and stress scenarios S1–S4, through both the credit and the liquidity channel.

    rsETH case · Would look healthy — HF 1.01–1.03. This is why one contour is not enough.

  • 03

    Liquidation feasibility

    LCR = liquidation bonus ÷ price impact, from a live aggregator quote. If liquidating is unprofitable, bad debt is likely.

    rsETH case · Selling unbacked collateral at size would not clear.

  • 04

    Behavioural anomalies

    Collateral inflow velocity, share of new addresses and of a single address, backing gap, oracle staleness, peg deviation.

    rsETH case · Sudden inflow of a single collateral from fresh addresses with a backing gap → gate and emergency response.

Principles

  • Fail-closedNo fresh score within its TTL means no allocation. An unavailable source invalidates the score; stale values are never served.
  • ProvenanceEvery axis is marked measured, default or not applicable. Two or more defaults trigger the NO_DATA gate.
  • ReproducibleA data snapshot plus a config version reproduces a decision bit for bit.
  • Audit trailEvery veto and every downgrade is logged with its rule.
  • AI without authorityAI classifies, watches and explains. It does not make allocation decisions — asset classifications stay pending review until a person confirms them.

Published

  • + Structure of the four risk contours
  • + Names of the seven score axes
  • + The eight hard gates, without thresholds
  • + The allocation ladder, including the 40% per-protocol cap
  • + Update frequencies

Not published

  • − Axis weights
  • − Base scores by asset type
  • − Penalty sizes
  • − Normalisation and anomaly thresholds

TBD What we publish — and what we don't — boundary pending confirmation