Restatement RiskSearchAskWhat happened (2021–23)How it works

How it works

  1. The app fetches public SEC data (XBRL company facts, filing index, comment-letter index, AAER list) with a declared User-Agent, under SEC's 10 requests/second limit, and caches it.
  2. For a report or a question, the model (Vultr Serverless Inference, glm-5.3-flash) writes a Python script.
  3. An internal runner service — the only component with the Docker socket, holding no keys — runs it in a fresh container: no network, data mounted read-only, 512 MB, 1 CPU, 64 processes, 30 s, no capabilities, non-root, empty environment. The container is removed after each attempt.
  4. Errors go back to the model, up to 3 attempts. Every attempt is shown.
  5. For reports, the model's numbers are checked against fixed reference code. Mismatches are shown; reference numbers are used.
  6. The model explains the result with the caveats below.
The SEC does not "audit" companies. Scrutiny comes as Corp Fin filing review (comment letters) and, rarely, an Enforcement inquiry. M-score (Beneish 1999) and F-score (Dechow et al. 2011) are published academic screens approximated on SEC XBRL data — not the SEC's non-public Accounting Quality Model. A screen is not a prediction.