Why this exists
A fund administrator's output gets signed by a person, and that person carries the liability. An agent that produces a 95% correct NAV pack is worth less than nothing, because checking the output costs more than doing the work by hand. So the questions that matter are: which numbers are computed in code, which have to come from a model, how a wrong one is caught before release, and what a person needs to see before they will put their name on it.
Both funds had the same shape of problem. Small team, a complex LLC agreement, quarterly statements rounded to thousands while the notes are in whole dollars, and a capital call calendar that depends on business days in a specific jurisdiction. Everything was in spreadsheets, and every quarter someone rebuilt the waterfall from memory.
Deterministic versus model
Extraction and judgment go to a language model. Arithmetic, business-day calendars and allocations go to deterministic Python. The split is explicit per stage and it is the product decision I am proudest of here.
| Stage | Model does | Code does |
|---|---|---|
| Terms extraction | Finds the fee and waterfall clauses, returns value plus clause and page reference | Nothing. Output is frozen to JSON and committed. |
| Statement tie-out | Nothing | 63 checks, each with an explicit tolerance and a source page |
| Fee recompute | Nothing | Rebuilds fees from terms over a NAV path, compares to reported |
| Capital call | Reads the drawdown instruction: purpose and amount | Business-day arithmetic, three allocation keys, equalisation, default handling |
| Waterfall | Reads the capital account history | Tier allocation per member, cap application, member offsets |
| Gate | Nothing | Enforces approver, title, decision, manifest state, artefact digest |
Extraction output is committed as JSON rather than computed at runtime. That is deliberate: the arithmetic is reproducible without a model, and the model's contribution is inspectable and diffable like any other input. The whole demo runs in under a second with no API key and no network.
What the run found
- The incentive cap. The agreement caps the manager's incentive at 20% of net gains after realised losses and unrealised depreciation. The portfolio carries $9M of unrealised appreciation, which does not count. A generic waterfall pays $1,408,810 and an auditor finds it.
- Three allocation keys. The agreement allocates a capital call differently depending on what the money is for. Splitting on commitment percentage alone misallocates every member, and hits hardest the one who has already defaulted.
- Columbus Day. 9 October looks like ten business days from 25 September. It is six, because the Fed is closed. The earliest compliant date is 16 October. This is the kind of thing that becomes a dispute with an investor's lawyer.
- Rounding is a product decision. Eight of sixty-three tie-out checks land inside published rounding and none break. Call both failures and the tool is useless. Call neither and it is dangerous.
The gate
Every stage writes its manifest with status: DRAFT and drafts its release artefacts without sending anything. Promotion requires a named approver, a title and an explicit decision. The gate writes that record into the manifest and refuses anything else. There is no flag to skip it, and a test asserts that a run with no approver cannot produce a released artefact.
release(
out_dir="outputs",
stage="capital_call",
reviewer="J. Okafor, Senior Fund Accountant",
decision="approve",
note="Exceptions E1-E4 accepted as proposed.",
)
Approval re-derives the numbers from the recorded inputs before emitting anything. If the draft no longer reproduces, the gate refuses to release it, so an approver is never signing figures that have gone stale under them. The record carries the approver's name, title, timestamp and a SHA-256 of the artefacts that were on screen.
An approval that cannot be attributed is not an approval.
What it does not do
I wrote the limits into the README because they are the next product decisions, not footnotes.
- There is no automated judge, so nothing yet checks the checker. Judge calibration against human reviewers is the next problem.
- The 63 tie-out checks were written against two statement periods. Nothing tracks what happens when a fund changes auditors and the note structure moves.
- Terms JSON is frozen by hand. If a fund amends its agreement, re-extraction is manual and there is no version diff.
- Everything is per fund, per period. Fleet oversight across hundreds of agents is out of scope.
The code, the sample run and the findings are public: github.com/solobuilderhub/fund-admin-agents. Built against a public drawdown fund's LLC agreement and a listed private equity trust's audited statements, with synthetic investor data. No client documents, no client names.