Every yield product creates a back office nobody built

Consider what actually happens when an institution takes a yield position onchain. A custodian holds the keys. A staking provider executes the transaction. The protocol records the lock, the cycle, the reward, the unlock. Each is doing its job. But ask the questions an auditor or fund administrator will ask — who owns this position, what has it earned, when does it mature, where is the evidence — and the answer is scattered across chains, custodians, contracts and a shared spreadsheet.

That missing record is the back office nobody built. It is the reason onchain yield has been hard to operate at scale, not the volatility or the custody — those are solved, or solvable. The unsolved problem is that the position never reaches the systems of record.

There is a well-understood category that solves exactly this for lending. Loan-servicing software does not originate loans and does not hold the assets. It services them: it tracks every loan through its lifecycle, computes what is owed, works exceptions, and lands the result in the systems the operator already runs. Lending has had this primitive for decades.

Onchain yield has not. Caddy is the servicing layer for onchain yield: what loan-servicing software is to lending, Caddy is to yield. It does not originate positions and it does not hold the assets. It services them — turning every onchain yield position into a verified institutional record with full evidence.

The layer applies the same five functions to every position, regardless of the underlying protocol:

• Model — agents interpret protocol mechanics and build the position adapter: the states, the entitlements, the lifecycle events.
• Track — follow every position across chains, custody records and offchain commercial terms, continuously.
• Investigate — agents work active exceptions end to end: gather evidence, propose the resolution, route for approval.
• Verify — deterministic code recomputes every financial result. No figure ships on a model’s say-so.
• Deliver — land the verified record in the custody, accounting and administration systems the operator already runs.

Each yield protocol becomes a position adapter: built once, verified continuously, reused by every operator holding that position. The cost of onboarding the second operator is a fraction of the first; the cost of the second protocol is a new adapter through the same five functions, not a new back office.

The trust story matters, because the obvious question for any AI-assisted financial system is: can I rely on the numbers? The answer is architectural. Caddy splits the work in two.

Agents do the interpretive work — reading protocol mechanics, modelling a position, investigating an exception, proposing a resolution. Deterministic code recomputes every financial result. No figure ships on a model’s say-so: the agent proposes, the code recomputes, and only the recomputed result is delivered. The whole layer is read-only by design — it never holds client funds and never signs a transaction.

That split is why a delivered record can carry an evidence pack that survives audit. Every entitlement, every lifecycle event, every resolved exception is source-linked and reproducible, not asserted.

What actually lands in the operator’s systems is a single, verifiable artifact — the position record:

• Identity — a stable position id, the protocol, the ownership (fund-segregated and attributed).
• Entitlements — what has been earned and accrued, per cycle.
• Lifecycle — locked, active, maturity, unlock.
• Exceptions — investigated, resolved and approved.
• Evidence — a source-linked, auditor-ready evidence pack.

The record’s status is not a claim; it is a recomputed, source-linked result: verified, and delivered.

The institutional adoption of onchain yield is no longer a question of whether but of how — and the “how” is gated by the back office. Custodians, asset managers, fund administrators, banks, neo-banks and corporate treasuries that need a verified, evidence-backed record of every position cannot build a new one per protocol. The category needs a servicing layer the way lending needed loan-servicing software: a primitive that is built once and reused across every operator and every protocol.

Native Bitcoin staking is one such adapter — lock state, cycle rewards, maturity and unlocks, modelled end to end. Any new yield product, from LSTs and restaking to basis trades and structured vaults, becomes an adapter through the same five functions.