> ## Documentation Index
> Fetch the complete documentation index at: https://docs.atlasyield.club/llms.txt
> Use this file to discover all available pages before exploring further.

# How the Engine works

> A mean-variance optimizer in the tradition of Markowitz — five deterministic steps.

A score rates one vault at a time. Turning scored vaults into a portfolio is a second,
separate job, done by the Atlas Engine — a mean-variance optimizer in the tradition of
Harry Markowitz. Markowitz, not magic: deterministic quantitative math, with its policy
parameters [published in full](/engine/tier-policy).

## The pipeline — five deterministic steps

<Steps>
  <Step title="Eligible universe">
    Start from every scored vault, keep only those that clear the chosen tier's hard
    gates — minimum Atlas Score, minimum TVL, minimum audit count and track record, an
    allowed asset class, and a permitted withdrawal speed. A vault that misses any single
    gate is never considered.
  </Step>

  <Step title="Return and risk inputs">
    For each survivor the Engine reads two numbers from public data: current APY as
    expected return, and the historical variance of that APY — extended to a covariance
    across vaults, with statistical shrinkage to steady thin histories — as risk. No
    private inputs, no forecasts.
  </Step>

  <Step title="Mean-variance optimization">
    Solve for the mix of vaults with the best expected return for the risk taken. A
    single risk-aversion dial, λ, set by the user's risk score, decides how hard the mix
    leans toward yield versus away from volatility.
  </Step>

  <Step title="Policy constraints">
    The raw optimum is held inside the tier's guardrails — caps per vault, per protocol,
    and per chain, a floor under each position, and a tighter cap on vaults with little
    history. Concentration is trimmed proportionally until every cap holds.
  </Step>

  <Step title="Final allocation">
    A set of positions with explicit percentages that sum to the full amount. Same inputs
    in, same portfolio out — every time.
  </Step>
</Steps>

<Note>
  When a candidate set has too little history for the covariance step to be trustworthy,
  the Engine falls back to a simpler, equally deterministic diversification rule rather
  than optimize on noise — and the result is labelled when that happens.
</Note>