An agent writes in one voice. A measured level, a headline and a stranger’s price target all come out in the same calm sentences. If a prediction goes in, a prediction comes out, and it now carries the authority of everything around it. The simplest defense is to keep predictions out of the input.
How a forecast gets in
There are three usual routes. You paste one in, as part of a note or a screenshot. A tool returns one, because the data vendor publishes targets or signals. Or your own instructions ask for one, with a line such as tell me where SPY is going. Each route ends the same way. The agent now has a forecast to pass on.
Why the agent cannot filter it
A model has no independent view of tomorrow. It cannot test a forecast against anything. What it can do is restate, and it restates well. A weak guess goes in and comes out with tidy reasoning attached. That added polish is the danger. The broader limits are set out in can AI predict the stock market.
What a reading is
A reading describes something that can be checked now. The call wall is at this strike. Funding is this rate. The index perp is this far from its level at the cash close. You can open a screen and confirm each one. A forecast cannot be confirmed until later, and by then the agent’s sentence has already been acted on.
How the tools hold the line
NoVo’s server is built to return readings only. The move on an index perp over a weekend is reported as a market price on another venue. get_base_rates describes its output as aggregate and never a forecast, and each cell carries its sample size. read_crypto_regime speaks to magnitude and location, never direction, and carries a field named not_a_signal. get_session_history describes what happened and does not predict the next session.
These labels exist for the model. A historical count is the kind of input a model most wants to turn into a prediction. The label tells it not to.
Your side of it
Keep your instructions clean. Ask what the readings are, not where price is going. If you paste in someone’s view, mark it as a view and ask the agent to keep it apart from the data. If the agent offers a forecast unasked, ask which tool returned it. No tool did.
The same test works on any answer. A sentence about the present should trace to a payload. A sentence about the future traces to nothing. This is the habit behind an answer you cannot check is an opinion.
What is left
Plenty. Where positioning sits, how stretched volatility is, what is scheduled, what holders are paying in funding. A trader can do a great deal with an accurate picture of the present. That picture is what the MCP & API serves, and what Dr. NoVo, the Financial Markets Super Intelligence, reads from when he is asked.