A hallucination is when an AI model produces fluent, confident output that is factually wrong or unsupported by any real source — a particular danger for financial data, where an invented number looks just as authoritative as a real one.
Hallucination arises because a language model predicts plausible text, not verified truth. Asked for a fund's position or a stock's insider activity from memory, it may generate a specific, wrong figure — and present it with full confidence.
The fix is architectural, not a bigger model: force the model to retrieve and ground its answer in live, sourced data, so it reports what a tool returned instead of what sounds right.
Asked "how many shares of AAPL does this fund hold," an ungrounded model may confidently state a precise number that is entirely fabricated.
Arkolith exists to stop this for market data: give the agent a live, sourced tool and a fabricated number becomes a real, citable one.
Arkolith turns this into live, sourced data your agent can query — SEC filings, insider activity, and market data behind one key, every datapoint traceable to its origin.