About this prototype
Independent speculative prototype: synthetic Atlas data, or public EDGAR data (not an official SEC product). “Atlas” is a fictional test corpus. It is not affiliated with, commissioned by, or endorsed by any government agency or company.
EDGAR data source (review option). A second, switchable corpus uses public EDGAR data: a dated snapshot (2026-10-31) of 29 companies’ recent filings and annual figures from XBRL company facts, fetched with a declared contact as SEC’s fair-access policy requires. It is an independent prototype, not an official SEC product, and it uses no SEC seal or branding. Form and term summaries are written for this prototype and link to the official source. An opt-in live mode fetches named companies on submit (never while typing) and falls back to the snapshot, visibly.
The question
Can an Ask-first, fluid composition interface help people state complex information needs more easily than a search field or a static AI prompt, without slowing down simple lookups?
What is real and what is mocked
- Real: the live interpretation. A deterministic rules classifier runs in your browser on every keystroke, with no network requests.
- Real: the native HTML form. It submits and shows results with JavaScript disabled.
- Real: editable comparison and timeframe controls with explicit-override semantics.
- Mocked (Atlas): every record and article is synthetic. Results are deterministic lookups, not AI-generated answers.
- Real data (EDGAR option): company filings and reported annual figures from a dated public snapshot, selected by documented rules (most recent filing wins for restated years; the XBRL concept used is shown).
- Not built: retrieval at scale, any language model, accounts, analytics.
Privacy
- Unfinished text never leaves your browser.
- A submitted question is kept in a short-lived (1 hour), httpOnly cookie in your browser only, so both result flows work without putting it in the URL. It is not logged or stored on a server.
- In EDGAR live mode the server requests the named companies’ public data from data.sec.gov; your question itself is not sent.
- A per-tab draft is kept in session storage so a failed submission doesn’t lose your text.
Limitations
- The rules recognize a fixed vocabulary; misspellings (“compair”) and novel phrasing are often missed. They fall back to calm, generic feedback rather than guessing.
- Interpretation labels are a heuristic. They are not calibrated probabilities and are not shown as percentages.
- No usability study has been run yet, so no claims of improved outcomes are made.