Methodology
How A Filing Becomes An Entry?
Every claim on this site quotes the filing it came from. Every score adds up from parts you can see. Below is the path a filing takes, from the SEC index to a published entry.
Evidence gate
Data Accuracy and Verification
Every claim must carry a verbatim quote from the filing. The evidence gate matches each quote against the source text and drops the claims it cannot find. We compute comparisons and percentage deltas in code, from SEC XBRL data.
“The company will restate its previously issued financial statement.”
“management expects the impact to be immaterial”
Score arithmetic
Materiality and Scoring
We score credit and equity impact by adding up weighted reasons. Each reason carries a direction, a weight, and a verified quote. Hard flags such as going concern or non-reliance cap the total, so a bad event cannot read as good.
| Reason | Dir | Weight |
|---|---|---|
| Non-Reliance on Prior Statements | neg | −2.0 |
| Covenant Waiver Obtained | pos | +0.5 |
| Auditor Resignation | neg | −1.5 |
| Hard Flag Cap Applied | −3.0 | |
Overview
A filing passes through six states, from the SEC index to publication. Item codes decide whether it gets one extraction pass or two. We commit each stage before the next, so an interrupted run resumes where it stopped.
- 01Discovered
- 02Pending
- 03Fetched
- 04Cleaned
- 05Extracted
- 06Published
-
Index Reading
The scraper reads SEC quarterly index files to find every 8-K in the target date range.
Source: SEC Quarterly Index
-
Accession Registration
The scraper registers each filing under its accession number, the permanent key that blocks duplicate processing. The row lands in the database as pending, carrying company, filing date, and form type.
→ Pending · One Accession, One Run
-
Item Resolution and Document Fetch
The ingester resolves item codes from local bulk archives, or asks the SEC directly when the archive misses. Triage then sets extraction depth, and the full submission comes down from SEC servers.
Pending → Fetched
-
Deterministic Normalization
The cleaner picks the most relevant document and strips its HTML down to plain text. Sentence boundaries survive and tables become pipe-delimited rows. We store that text verbatim, because the evidence gate matches quotes against it character for character.
Fetched → Cleaned · Same Input, Same Output
-
Model Calls, Voting, Evidence Gating
Cleaned text goes to the model under a schema that demands a quote for every claim. Normal filings get one call. Severe ones get two, merged by majority vote on the discrete fields. We validate the output against the schema, then send it through the gate.
Normal → 1 CallSevere → 2 Calls, Majority VoteCleaned → Extracted
-
Commit and Publish
We commit the raw document and the cleaned text as we go, so nothing has to be fetched twice. Once an extraction clears the gate, the event goes to the index.
Extracted → Published · Resumes At Last Stage
Known Limitations
- SEC filings and XBRL coverage vary by issuer and form, so prior-period comparisons may be unavailable.
- Event classifications and extracted figures are research outputs, not issuer-authored labels.
- Amended filings must be assessed against their underlying filing by the reader.
- 8K / Observer is not investment advice and should not replace direct review of the underlying EDGAR documents.