How a figure gets on the record.
Every figure in Tark is traceable to a public filing or report. Our methodology shows how we source the data, calculate results, and assess our final benchmarks.
- Methodology
- The pipeline
- Evidence labels
- Six-factor review
- Benchmark engine
- Computed figures
- Commitments
The same six steps for every document
Collect
The pipeline asks each publisher what is new: SEC EDGAR, Form ADV and Form D, the DOL’s Form 5500 file, sponsor sites and — public LP reports. A publisher that is down is passed over and listed.
Read
Figures come from tagged filing data where it exists: N-PORT, N-CEN, XBRL, Form ADV and the Form 5500 schedules. A value that exists only in prose is taken with the sentence it came from.
Read again
Rows parsed from a PDF’s text layer are read a second time by OCR from the rendered page; the Sources page shows, for each public LP, how many of its rows have been. The parsed numbers must appear in that read. A disagreement is flagged beside the row. OCR text never becomes a figure.
Label
Each value is labeled source-linked, computed, unverified or inferred, and each gap is named: not found publicly, or not applicable to the wrapper. Source-linked means machine read from tagged filing data and linked to its filing, not confirmed by a person. An unverified value stays marked until a reviewer confirms the quoted text. A gap is never estimated.
Test
Every derived table is rebuilt and the test suites run. A snapshot is published only if they pass. A failed test stops the run, and the site keeps the last published snapshot.
Keep
Source files are held as served. A file read from the Internet Archive is checked byte for byte: the SHA-1 of the bytes held must equal the digest the archive recorded.
Every value labeled, every gap named.
Every cell in a record carries a label, so a reader can tell what came from a structured filing and what did not, and what kind of gap an empty cell is.
- Source-linked
Machine read from a structured filing (tagged data such as N-PORT, N-CEN, XBRL or the Form 5500 schedules) and linked to it. No person has confirmed it.
✓ N-PORT · 2026-03-31 - Unverified
Read from prose. The supporting quote is kept in the record’s appendix until a reviewer confirms it.
? Shareholder letter, p. 3 - Inferred
Taken from an entity’s own facts, such as a K-1 reported on a partnership’s form.
Inferred · Entity’s own filing - Zero on file
Found, and the value is zero: the filing states nil, such as a Form D’s sales commissions of $0. A zero is a figure, not a gap.
0.00% · zero on file - Not found publicly
The field applies to this kind of vehicle but is not in the filings read. Tark does not record a failed read, so an omission and a value its reader missed look the same. Nothing is assumed in its place, and a gap is not a refusal to disclose.
Not found publicly - Not applicable
Fewer than one in two hundred vehicles of the kind carry the field on the public record. It is listed once and not counted for or against the fund.
Not applicable to this wrapper
Verified share of evidence cells, by wrapper
Cells read from a structured filing, over all cells for that wrapper.
Non-traded REITs report mostly in prose, so most of their cells wait for a reviewer.
Coverage is not a score. A percentile is not a grade.
Tark reviews a fund on six factors: performance, fees, liquidity, valuation, benchmarks and complexity. Methodology version —, stated in every record written under it.
- Evidence coverage“7 of the 10 applicable fee fields found”: how much of the fund’s public record is on file for a factor, counted over the fields that apply to its wrapper. Written in words, never as a fraction, and never a rank or a view of the fund.
- Evidence minimumWith fewer than — fields found, a factor is read field by field and no summary of it is drawn.
- Peer positionThe share of peers with a lower value: vehicles of the same kind and strategy, the same target year for a target-date fund, the same period for a period’s figure, and four or more carrying the figure. Each figure says what a higher value means, never whether higher is better.
- WeightingNone. No factor is scored, weighted or added into an overall score. The assessment is the reviewer’s, and the record keeps it in the reviewer’s own words beside the evidence, the gaps, the assumptions and the comparisons it was made on.
- Whose methodTark’s. The factor names follow the proposed rule cited below; whether a review on this method satisfies that proposal or any final rule has not been verified, and Tark does not say it does.
How the engine chooses or refuses.
Every fund is tested against up to — candidates in four lanes (the first, its own stated benchmark, only where the fund states one). Each is scored against the same rubric, and every candidate that loses keeps its reason.
Stated
The benchmark the fund names in its own filings.
Published indices
Indices in the registry, each with its publisher, owner, basis and licence.
Peer cohort
Funds with a similar strategy and valuation profile, built from filings. Shown beside, never as the benchmark.
PME and Direct Alpha
Public market equivalent math against a public index, where its series is held.
— candidates a fund · —
What the rubric weighs
| Criterion | What it weighs |
|---|---|
| Independence | Distance between the index publisher and the fund’s adviser. The adviser’s own index never fills the slot. |
| Strategy | Fit between the index and the fund’s filed strategy profile. |
| Publication | Whether the index publishes its rules, and whether it can be invested in. |
| Liquidity | Fit between how the index is valued and how the fund is valued and redeemed. |
| Data held | Whether the record holds the index’s level series. |
| Self-stated | Whether the fund names the index in its own filings. |
Criterion summaries are plain-language. The engine specification governs.
Rules
- Clearing the bar fills the slotThe adviser’s own index never does.
- Ties break on independenceThe tie is recorded as tied on the rubric’s score.
- Descriptor onlyIf the series is not held, the index is named and no number is shown.
- The engine can refuseWhen nothing clears the bar, the record says so.
Dispositions
Arithmetic on filed numbers, else modeled and flagged.
Every computed figure can be re-derived from the filings it cites. A model output is shown beside the filed figure and marked as a model.
BComputed figures
Definition B.1 (Return). For filed monthly returns r1, r2, …, the return over a period of months t compounds as
Series of different frequency meet on the coarsest calendar they share.
Definition B.2 (De-smoothed return). Where the first-order autocorrelation ρ of a series is significant,
and σ* denotes the volatility of r*. It is reported beside the filed volatility and marked as a model.
Definition B.3 (Sharpe ratio). With μ0 the mean filed return of a Treasury-bill fund,
Definition B.4 (Peer percentile). For a value x among peers x1, …, xn,
A position among peers, never a grade.
Definition B.5 (Gap to drawdown funds). With dk the return of a composite of drawdown funds in quarter k,
Definition B.6 (Liquidity coverage). With Q the repurchase capacity of the position, w its weight in the plan and O the plan’s outflow,
Definition B.7 (PME and Direct Alpha). Measured against a public index, where the record holds its series.
Definition B.8 (Index proxy). An index-tracking fund’s filed return may stand in for an index the record does not hold. It is labeled as that fund’s return, never as the index.
Proposition B.9. Every figure in Definitions B.1 to B.8 is computed only from values the record cites.
Proof. By construction. No missing input is estimated, carried forward or assumed.□
Oct 2025: r* = (0.48 − 0.65 × 1.06) / 0.35 = −0.60%
| rt | Filed monthly returns, Apr 2025 to Mar 2026 | 12 ✓ |
| ρ | First-order autocorrelation of the filed series | 0.65 ✓ |
| σ | Volatility of the filed returns, a year | 0.99% ✓ |
Illustrative. Modeled on a real record, anonymized.
What Tark will not do.
Five lines the product does not cross. Each one is enforced in how the record is built, not left to a reader’s judgment.
- 01
Recommend, approve or select an investment.
The committee decides. The record documents how it decided.
- 02
Show a figure without its source.
Every value names its filing and section. A value without one is not shown; the field reads not found publicly.
- 03
Estimate what a filing does not state.
A gap stays a gap. Nothing is carried forward, averaged or modeled into it.
- 04
Substitute one index for another.
If an index series is not held, the record names the index and shows no number.
- 05
Let OCR text become a figure.
The second read can only flag a row. The parsed value stands until someone looks.
U.S. Department of Labor, proposed rule, 91 FR 16088 (RIN 1210-AC38).
Put your own lineup on the record.
Bring the alternatives you are weighing, and we will build their records with you. Let us help you compare apples to apples, and oranges to oranges.