Known limits

Datagoat isn't the right tool for every question. These are the edges we know about, so you can design around them.

What it can't answer

Anything without history. Datagoat learns from outcomes you've recorded. A record needs about 500 labeled rows. Below that, the answer is not_yet. For a new question with no history, use a zero-shot model like Jev.
Free text. Text columns count as categories, not as language. To use what a message says, turn it into columns first (Read, then learn).
Numbers. Outcomes are yes/no. "How much will they spend?" needs a regression tool. "Will they spend over $500?" works, if the record has that column.
Causes. Reasons and choice picks are associations in the record. Switching a customer to the picked contract may not keep them. Report what happens, and let the outcomes tell you.

A refusal stays a refusal. The same record and the same question always refuse the same way. Retrying is pointless. Change the record (more rows, more columns) or the question.
The first fit takes time. Seconds on a small table. A minute or two on a large record, and then the call returns pending. A repeat call on the same bytes answers from the cached fit.
One changed byte is a new record. Appending a day of rows makes a new fit. Use refit_of to link the two, and dg_drift to see whether anything moved.
outcome_is_desirable is part of the fit. Two questions on one outcome share a fit only when they say the same thing about it. One with false and one with nothing set run two fits.
rank beside other types ranks only the cases. The other types need cases, and rank ranks what it's given. Rank the whole record in its own call.
Choice needs the option to matter. When the option column isn't part of the pattern, the answer is refused (option_not_in_pattern), not a tie-break.

Levers appear only where the engine can show the changed state is reachable on held-out rows. Many answers carry none, and then there is nothing to attest.
dg_evidence compares groups only once each has 30 cases with a reported outcome.
Acted cases are chosen by you, not at random. A difference is evidence, not proof of cause.
Models fitted before lever tokens were added carry levers without tokens until they refit.

Up to 8 questions and 10,000 cases per call. Inline records up to 10,000 rows (or 8 MB of CSV). Larger records go through dg_add_dataset.
rank returns at most 1,000 cases.
Over MCP, a large answer can overflow an assistant's context. Ask about the cases the conversation needs, and use the REST API for bulk scoring.

Shapes other than table answer about cases by id, not new rows.
The reading's windows and horizons come from fixed sets (for example, horizon_days of 30, 60 or 90). Other values are refused, with the allowed ones listed.
An event log shorter than its windows reads to nothing (shape_empty).

A Verdict expires with its model, 90 days after the fit, even when a later call reused that fit. After that, the question fits again.
Verdicts name datagoat.io as their issuer and verify against Datagoat's public keys.