Open the description of a NoVo tool and you will find sentences that read like warnings. Some are in capitals. They look odd in documentation. They are there because the reader is a model, and a model reads the description every time it considers the tool.
Why the caveat goes in the description
A person learns a data source once and remembers its quirks. A model starts fresh in every conversation. It knows only what the catalogue tells it. If a quirk lives on a documentation page the model never opens, the quirk does not exist for it. So the warning goes where the model must look. The general case is made in documentation is runtime.
Warnings about what a number is
Some warnings stop a label from being misread. The delayed quotes tool has rows named after indexes that are in fact futures, and the documentation says so, so an agent does not quote a future as the cash index. get_earnings_dates states that it passes through a public exchange calendar and is not a NoVo measurement. get_retail_chatter states that it counts tags users put on their own posts, on one venue.
Warnings about time
get_congress_trades returns House disclosures. Its description says these are filings, not live trades, and that members have up to 45 days to file. Without that sentence an agent will report a weeks-old disclosure as something a member just bought. With it, a careful agent says when the trade was made and when it was filed.
Warnings about gaps
get_crypto_depth warns that a null cell on the volatility surface is a gap in the book and not a zero. get_dealer_levels tells the model to read the gated field before deciding a number is missing. get_ticker_brief says its dealer section is not covered for any name outside SPY, QQQ and IWM. Each one blocks the same mistake, which is turning no data into a statement.
Warnings about what not to conclude
get_base_rates says its output is aggregate and never a forecast, and that a cell below its sample floor must not be quoted. read_crypto_regime carries a field named not_a_signal. get_session_history says it describes what happened and does not predict the next session. These exist because a model handed a historical count will reach for a prediction unless told otherwise.
How a model uses them
A good model does two things with a warning. It uses it to choose the right tool. And it carries the caveat into the answer, in its own words. You can test this. Ask about a congressional trade and see whether the filing lag appears. Ask for the IWM perp and see whether it says none exists, as in asking an agent where SPY opens.
Models do skip caveats, more often in long conversations. If yours does, remind it: repeat any warning in the tool’s description that applies to the figure. The warnings are public. They are listed with each tool on the MCP & API page, so you can check the agent against them.