Methodology

How each signal is produced, and what it does and does not mean.

How much is flagged

A count of events means little without the population it came from. Across the filing days processed so far:

Filings in the daily index

47,722

Read in full

3,864

Events published

251

Flagged

6.5%

Most filings are never candidates: fund and securitisation paperwork makes up the bulk of daily volume and cannot carry these signals. The flag rate is deliberately low — these are meant to be rare events.

Where the data comes from

Every weekday the pipeline reads the SEC's EDGAR daily index — the official list of everything filed that day. It keeps 8-K, 10-K, 10-Q, 20-F and 40-F filings from operating companies, discarding fund and securitisation paperwork, which makes up the majority of daily volume and cannot carry these signals.

Restatements and auditor changes

These come from the SEC's own structured item codes, read from each filing's header. Item 4.02 is "Non-Reliance on Previously Issued Financial Statements"; Item 4.01 is "Changes in Registrant's Certifying Accountant". Because the filer selects these codes themselves, a match is a fact about what was disclosed, not an inference. These are labelled SEC item code.

The sub-classification is the signal. The same item code covers materially different events, so the filing itself is read to separate them. Under Item 4.02(a) management or the board reached the conclusion; under 4.02(b) the auditor told them, which is more serious because the company did not find it itself. For Item 4.01, an auditor resigning is a stronger signal than a company dismissing one, a disclosed disagreement stronger still, and a move from a large firm to a much smaller one is ranked above a rotation between equals.

What Item 4.02 does not mean. It means the company told investors that earlier figures should no longer be relied upon. It is not by itself evidence of fraud or misconduct. Restatements frequently arise from technical accounting corrections.

Late filings

A company that cannot file on time is often the earliest public warning, and it usually precedes a restatement rather than following it. These come from Form 12b-25 (filed as NT 10-K or NT 10-Q).

Beyond the fact of lateness, the form asks whether the company anticipates a significant change in results of operations, answered as a checkbox. A "yes" is the company pre-announcing that the delayed numbers will look materially different, and those are marked elevated. So are late annual reports, and companies that are also behind on their other periodic reports.

Going concern

This is the signal most often done badly, so it is worth being precise.

Searching filings for the phrase "going concern" does not work. The phrase appears in the Risk Factors of most speculative issuers as standing boilerplate, unchanged for years. It also appears in forward-looking-statement disclaimers. None of that is news.

Instead the pipeline locates the actual accounting note — the disclosure made under ASC 205-40 — and reads its conclusion. That note always opens by reciting both possible outcomes as methodology, so the conclusion sentence, not nearby wording, decides the classification. Each filing is placed on a ladder:

1. No going-concern disclosure
2. Risk-factor language only — boilerplate, no accounting conclusion
3. Substantial doubt raised, alleviated by management's plans
4. Substantial doubt about the ability to continue as a going concern

An event is published only when a company moves between rungs, compared against its own previous filing of the same type. A company that has disclosed substantial doubt for eight consecutive quarters generates no event, because nothing changed.

Accounting policy and revenue recognition beta

The pipeline extracts the significant-accounting-policies note and the revenue-recognition note, strips out figures, dates and quarter labels, and compares the wording against the same company's previous comparable filing. Comparisons are always like-for-like — a 10-Q against a 10-Q — because a quarterly note is far shorter than an annual one and comparing across them would flag every company every year.

These are marked beta. They are derived from text comparison rather than a structured code, so they carry more noise. Treat them as a prompt to read the filing, not as a conclusion.

The role of AI

Detection is entirely deterministic: pattern matching, item codes and text comparison. No model decides whether something is a signal.

When enabled, a language model adds a plain-English summary and screens beta text-diff findings for cosmetic rewrites. Anything it writes is labelled AI summary and always appears alongside the verbatim filing text and a link to the source. If the analysis layer is off or fails, events still publish — with evidence and quotes intact.

Limitations

• Only filings from the days the pipeline has processed. It is not a historical database.
• Section extraction depends on filing structure. Unusual formatting can cause a note to be missed.
• Companies with no prior comparable filing produce no comparison signals: a change cannot be shown without a baseline.
• Amended filings are compared against the original series, which can surface wording differences that are procedural rather than substantive.

Corrections

Every entry links to its source. If an entry misreads a filing, open an issue on the project repository with the accession number and what it should say.