Plain-English definitions for the SEC, FINRA, and AI-research terminology used across this site. Every entry has a definition, a “why it matters”, and where useful, a worked example. Cross-linked back to the source documentation. Public, cite-able, no auth required.
The actual number of shares short on a given settlement date (T+2), reported by the self-regulatory organization twice per month for every US-listed equity.
▌ WHY IT MATTERS
Settled SI is the truth. Daily short volume is intraday flow — most of it flat by the close. When you say %-of-float-shorted, you mean settled.
▸ A mega-cap's settlement file shows ~140M shares short for the period.
A US-listed certificate representing ownership of foreign-domiciled shares held by a depositary bank. The ADR trades in USD on a US exchange while the underlying ordinary shares trade abroad.
▌ WHY IT MATTERS
ADRs let US investors hold BABA, TSM, NVO, etc. without trading the foreign primary listing. Market-cap math requires the deposit ratio to be applied or it will be off by 5×, 8×, or worse.
▸ Some foreign issuers package 8 ordinaries per receipt; others package 5; others 1. The ratio is always in the deposit agreement.
The number of ordinary (foreign) shares each ADR represents, established in the Form F-6 deposit agreement filed with the SEC.
▌ WHY IT MATTERS
SEC XBRL reports the foreign issuer's ordinary share count. To compute a correct ADS-equivalent market cap, divide ordinaries by the deposit ratio before multiplying by the US-listed ADR price.
▸ An issuer reporting ~18.6B ordinaries with a 1:8 deposit ratio has ~2.32B ADS-equivalents; multiply by the receipt price to get the correct market cap.
The SEC registration statement governing the issuance of ADRs by the depositary bank. F-6 = initial, F-6EF = becomes effective on filing, F-6 POS = post-effective amendment.
▌ WHY IT MATTERS
The deposit agreement language ("each ADS represents X ordinary shares") lives in the F-6. Parse it to extract the deposit ratio programmatically.
An LLM generation pattern that conditions output on chunks retrieved from a corpus at query time. The model translates the chunks; the chunks ground the truth.
▌ WHY IT MATTERS
RAG is how TAPE 33 prevents synthesis hallucination. Every AI-generated claim is conditioned on retrieved evidence — no chunk, no claim.
A high-dimensional vector representation of a text chunk, computed by a model like the embedding model (1536 dimensions). Used to find semantically similar chunks via cosine similarity.
▌ WHY IT MATTERS
Embeddings power retrieval. Two chunks with similar embeddings are likely about the same thing, even if they share no exact words.
Indexing + querying technique for finding the top-K most-similar embeddings. the embedded vector-search engine provides brute-force cosine similarity at sub-100ms for ~2K-chunk corpora.
▌ WHY IT MATTERS
VSS is the retrieval engine under the synthesis. The corpus is small enough that brute-force works; HNSW persistence currently breaks on re-index.
A US equity exchange known for its 350-microsecond speed bump (Tiingo's real-time data source).
▌ WHY IT MATTERS
IEX trades are real but not the full tape — a quote from IEX may differ from the SIP NBBO. Use as a real-time proxy with an explicit understanding of the gap.
federally-registered date by which a trial's primary endpoint will be assessed. The proxy for “data readout” when an explicit readout date isn't published.
▌ WHY IT MATTERS
Use as the catalyst date when no PDUFA exists. Note that companies routinely slip primary-completion dates without re-registering.