Ask a general-purpose chatbot what a company’s most recent 8-K said, or which insiders bought last week, and you will often get a clean, specific, completely made-up answer. The numbers look plausible. The tickers are real. The filing it cites may not exist.
Why this happens
A language model predicts text. With no tool to look something up, it fills the gap with the most likely-sounding sequence of words. For market data that is a real problem, because plausible and correct are not the same thing, and a wrong insider-buy figure can send someone down the wrong path with actual money on the line.
The common instinct is to wait for a bigger model. A bigger model still guesses when it has nothing to read. Model size does not change whether the thing has access to the actual filing.
The fix is boring: give it the data, with the source
When an assistant can call a tool that returns the real signal, the real filing, and a link back to where it came from, it stops guessing and starts reading. Because the source travels with the answer, a person can check it instead of trusting it.
That is the whole idea behind connecting models to data through something like the Model Context Protocol. The model asks for list_signals or sec_insider_trades, gets structured results, and answers from those instead of from memory.
The honest pushback
You could argue this just moves the trust problem: now you are trusting the data provider instead of the model. Fair. The difference is that a sourced answer can be audited. If a signal links to the Form 4 it came from, you can open the filing and look. A hallucinated answer links to nothing, because there was never anything there.
It also will not fix everything. A model can still misread good data, or an agent can ask the wrong question. Source-linked data raises the floor. It does not remove the need to check.
Where this leaves us
We build QuantConomy around one assumption: the same source-linked signals should be readable by a person and by an agent, over the same API and the same MCP server. Not because agents are the future and humans are the past, but because both are better off when the data they read is real and the source is one click away.