cache

Architecture

One vault,
two ways to use it.

cache. is an ERC-4626 vault on Base. You deposit DIEM; the vault stakes it on Venice and resells the daily inference allowance, compounding the proceeds back into DIEM. That is the sell-side path — the default, and the only one your deposit needs.

HYBRID mode adds a second path: instead of selling all of your share of the daily allowance, you can spend part of it yourself, as inference, for free.

How HYBRID inference works

Every depositor holds vDIEM — their ERC-4626 share of the vault. The vault has a fixed daily Venice allowance in USD. Your free inference quota for the day is simply your pro-rata slice of that allowance:

quotaUsd = (your vDIEM / total vDIEM) × vaultDailyAllowanceUsd

The split is measured at a fixed snapshot — the block at the start of each UTC day — so it can't be gamed intra-day. Whatever you don't spend stays in the sell-side path and still earns yield. Nothing about HYBRID mode changes your vDIEM or your ability to redeem.

The proxy, and what it can't do

  1. You sign one SIWE message to prove you hold vDIEM. In exchange the proxy issues a bearer API key. That is the only signature inference ever needs — every later request uses the key.
  2. The key is shown once and stored only in your browser, encrypted at rest with a key derived from a wallet signature. That encryption protects a copied browser-storage blob — it is not a secrets vault, so treat the key like a password. It never touches a server we run beyond the request it authorizes.
  3. The proxy mirrors Venice's API and meters each request against your remaining daily quota. It can refuse you past your quota; it cannot move your DIEM, mint or burn vDIEM, or touch the vault.
  4. Usage is committed weekly as an on-chain Merkle root, with the leaves published to IPFS — so you can independently verify the bytes the proxy charged you for.

The full picture

The complete system — the autonomous Smart Account that harvests and settles, the AntSeed seller delegation, the 1Shot gas relayer, and the end-to-end sequence diagrams — is documented as mermaid diagrams in the repository:

agent/cache-architecture.md

That file is the canonical reference; this page is the depositor-facing summary the inference flow links to.

Simplicity at the front, machinery at the back.
cache