Incentive Fee Terms Tracker — 2025-26 annual reports

Carry mechanics are stated numerically in almost no primary public document. A private fund's limited partnership agreement is not filed, and the carry and compensation figures in general circulation trace back to recruiter surveys and vendor blogs whose sample, method and field year usually cannot be checked. This tracker exists because there is one exception, and this is it.

A business development company is a closed-end fund that has elected regulation under the Investment Company Act. Its adviser's compensation is subject to a statutory disclosure and approval regime, so its annual report prints the incentive fee rate on income, the hurdle, the catch-up, the capital gains fee and any look-back in plain numbers — inside a document with an accession number that anyone can open.

Every company with a Form N-54A business development company election on EDGAR that filed a Form 10-K in the window, one filing per company — the most recent 10-K in the window. Membership is structural: 365 companies have a Form N-54A election on file in EDGAR full-text coverage, and 177 of them filed a 10-K between 2025-07-01 and 2026-06-30. 188 companies were excluded, each with its reason recorded in the dataset — most of them because they filed no annual report in the window, and some because they had withdrawn the election on Form N-54C.

Of the 177 funds in the universe, 170 are externally managed and pay an adviser, and 7 are internally managed and pay no advisory or incentive fee at all. The internally managed funds stay in the universe and are reported as a finding. They are excluded only from the fee denominators, because a fund with no adviser is not a fund that declined to disclose its adviser's rate.

EDGAR full-text search reaches back to 2001. A company that elected before 2001 and never re-elected is therefore outside the election frame. That is a stated boundary of this universe rather than a silent gap, and it is the one place where the frame is narrower than the population.

Each annual report is converted to plain text and read by pattern parsers that record both the value and the sentence it came from. The stored excerpt is anchored on the number rather than on the start of the pattern, so every figure in the dataset can be checked against the sentence printed beside it without opening the filing.

Nothing is computed. A quarterly hurdle is not multiplied by four to produce an annual one; where a filing prints only the quarterly rate, the annual field stays empty. Where a filing states only a change — an agreement amended to reduce a fee from one rate to another — the new rate is taken and the old one is never read as current.

These are keyword-and-shape parsers, not a legal reading of an advisory agreement. A term described in unusual language records as absent, so every presence rate here is a floor and never a ceiling.

Four traps were found by reading parser output against the filings, and each now has a fixture test built from the real filing text. The capital gains fee rate sits one clause from the income fee rate and reads as it if nothing separates them. A quarterly hurdle of 1.75% and its annualized 7.00% are each other's most available wrong answer, so a stated annual rate that does not reconcile to four times the stated quarterly rate is discarded rather than published.

A fee rate recited from a superseded agreement looks exactly like a current one and sits in the same section, so a candidate whose own sentence carries an amendment date or pro-forma language is thrown away. And every prospectus carries a worked example with invented income figures printed beside the real hurdle — read from there, this tracker would publish an assumed return as a negotiated term, so worked-example language disqualifies a candidate outright.

Two more are structural rather than textual. An internally managed BDC that owns an asset manager describes in its own annual report the fee schedule it CHARGES other funds; the parser detects internal management first and never reads a fee for such a fund. And a risk factor noting that an adviser "may earn an incentive fee even in a quarter in which we incur a loss" describes the absence of a look-back, so presence is read only from a defined term in the agreement and never from a bare mention.

The pipeline is scripted end to end: build the election frame, select each company's annual report, fetch, parse, aggregate. Search results, per-company filing selections, XBRL total-asset facts, and per-filing extracts with their matched sentences are all cached in the repository, so the tables rebuild offline from committed inputs.

One command regenerates everything from EDGAR: `npx tsx scripts/carry-terms/run.ts --refresh`. Run without the flag to rebuild from cache. Requests carry a descriptive User-Agent with a working contact address and run at four per second against the SEC's published ceiling of ten, with exponential backoff. Every endpoint used is free and public; no paid data source is involved.