How the filings are read

Printed and computed

A printed figure is the filing's number under the filing's own label, with the page it is on and a link to the filing on EDGAR. A figure the filing does not print says "not printed in the filing", with the reason where we know it: the public copy leaves it out, the firm computes net capital under the alternative method, or the filing prints only its formula. A field we have not read yet says "not yet read". Nothing is left blank.

Anything we computed says computed, and in the report shows its formula in the filing's own figures. The computed ratios:

Net capital as a multiple of requirement

Net capital ÷ the minimum net capital requirement, so 1.0× is the floor. Where the filing prints no requirement, net capital ÷ (net capital − excess net capital), only where the figures show that difference is the Rule 15c3‑1(a) requirement, and the page says so.

Aggregate indebtedness to net capital

Aggregate indebtedness ÷ net capital, where the filing prints aggregate indebtedness in dollars and net capital is positive.

Equity to total assets

Equity ÷ total assets.

Equity not counted in net capital

(Equity + subordinated borrowings − net capital) ÷ equity, where equity is positive.

Margin

Net income ÷ total revenue, where the filing includes an income statement.

Which filings are read

Every broker-dealer with an annual report on Form X-17A-5 on EDGAR, from filings through September 15, 2026: 3,061 filers, of which 3,040 have a page. A firm that has never filed on EDGAR is not among them. Each page reads one filing: the firm's latest fiscal year, and in it the last filing EDGAR took that carries a statement of financial condition.

A further 3 filings state their figures in another currency. Their pages show identity, auditor, opinion and flags but no figures, and they sit out of bands, placements and counts.

What the public copy leaves out

Rule 17a‑5(e)(3) makes the statement of financial condition, its notes and the auditor's report on it public. A firm may bind the rest of the annual report separately and ask for it to be kept confidential: the income statement, the cash flows and the net capital computation are public only where the firm did not. The quarterly FOCUS report is confidential.

So revenue, net income and margin describe the firms that publish them, most of them small. Each band's page compares the median size and cushion of firms that publish the income statement with those that don't.

How often the reading is wrong

A blind second read of 2%

63 filings from the reading the site shows, read again blind, by a procedure aimed at misreads that still add up, every disagreement ruled on by a referee (September 25, 2026). 0 of the 63 had an error in a figure or a flag: 0 of 630 money figures, 0 of 378 other figures, 0 of 315 flag and classification fields. At 95% confidence, at most 4.6% of filings carry one. Page citations: 4 of 583 wrong, all in one filing, since corrected. One filing from each of the long runs, read the same way (September 28, 2026): 0 errors.

Flag sentences against the page

All 207 quoted going-concern, material-weakness and net-capital sentences and shortfalls, read against their page images by a second model (September 25, 2026). 0 of 106 going-concern quotes, 0 of 6 material-weakness quotes and 0 of 27 shortfalls were wrong. 6 sentences about a cure or management's plans were wrong, and have been corrected.

Against an earlier study

100 filings read against an earlier study of the same filings (September 17, 2026): 976 of 1,000 money figures matched outright, and of the 24 that didn't, a blind referee sided with our reading on 21, with the study on 2, and with both on 1. No digit was misread, and all 20 known defects in the documents were caught.

Two reads of the same filings

1,727 filings read twice (the first reading against the one the site shows, September 24, 2026): 61 of 17,270 money figures moved, none of them total assets, equity, net capital or net income. How a firm clears moved on 6 to 10 of every 100 filings, mostly to or from not stated.

For anyone checking our records, the runs: the site shows corpus-2026-v2, with corpus-2026-long and corpus-2026-split for the longest filings; the first reading was corpus-2026; the blind second read, audit-2026-v2; the earlier study, n100-sonnet.

Not measured: expense sharing, the labels as printed, and a misread the second read shares with the first, since the same model first read most of the audited filings. The audit reads the same text layer as the record, so an error that starts there can pass both.

The facing page and the document

EDGAR carries each filing's facing page as data: the filer's own entries for the period, the auditor, and whether the report finds a material weakness. Where the facing page and the document disagree, the firm's page prints what the document says. A firm on the SEC's list of active broker-dealers shows the list's file number. The auditor is whoever signed the opinion; where the public copy has no auditor's report, it is the accountant the facing page names, linked to the facing page.

The facing page's material-weakness answer disagreed with the document in 105 of the 3,119 filings the site reads (September 28, 2026). In 104 the facing page answers yes and the document reports none; 70 of those firms file an exemption report, which carries no finding either way.

The early-warning counts

/peers counts the firms at Rule 17a‑11's lines, naming none: aggregate indebtedness above 12 to 1, net capital under 120% of the requirement, and net capital below the requirement. Each count reads one day, each firm's fiscal year end, and says nothing about any firm since.

The aggregate-indebtedness ratio counted is the one placements use: aggregate indebtedness printed in dollars over positive net capital. A filing that prints the ratio but not the dollars is left out. A printed ratio above 12 with no dollars behind it is most often a percentage printed as "to 1", implying indebtedness dozens of times the firm's total liabilities, and indebtedness worked back from the requirement fails the same test.

What is free elsewhere