Glossary
Every Datagoat term once, with where to read more and, where one exists, its Jev counterpart.
| Term | Means | More | Jev |
|---|---|---|---|
record (data) |
past cases and their outcomes, the evidence Datagoat learns from; a stored record is a dataset | The record | state (the input) |
| case | one thing a question is about: a customer, a deal, a machine, a run | The record | the state itself |
| entity column | the column naming each case (entity_column) |
The record | none |
| outcome column | the yes/no column a question learns | Questions | instructions |
outcome_is_desirable |
whether the outcome is one you want; Datagoat never fills it in | Questions | none |
| question, question id | a typed question (yesno, score, choice, rank) under an id you choose |
Questions | question, id |
| answer | one per question id, with a state |
Answers | answer |
state |
answered, refused or not_yet: the answer's status |
The answer state | confidence |
| refusal | the record holds no pattern that held up on unseen rows; the same call returns the same refusal | The answer state | none |
not_yet |
too few labeled rows to check a pattern | Known limits | none |
p |
the chance the outcome happens for a case | Answers | Noul probability |
| reason | a column that moved a case's chance, its value, the direction and the range its value fell in |
Reasons | none |
pattern, pattern_match |
where the outcome is most common, with cut points; how many of those conditions a case meets | The pattern | none |
| band | the engine's gate per case: act, escalate or refuse, with band_reason and max_autonomy |
The band | none |
| Verdict | the signed record of an answer: model, record hash, cases and scores | Verdicts | none |
| fit | learning a model from a record, checked on held-out rows; the first ask on a record runs one | Answers | none |
model_ref |
the id of a fitted model; a question can name it to score with no fit | Score from the model | none |
| namespace | a separate set of models and usage inside a workspace, such as one per customer | One namespace per customer | none |
| shape | how a record that is not one row per case is read: events, series, panel, signals, traces, snapshots | Shapes | none |
| held out | rows the model did not learn from, used to check it | How answers are checked | none |
| top-decile lift | on held-out rows, how many times as often the top tenth by p had the outcome as average |
Quality | none |
lever, lever_token |
a feature whose change is associated with a lower (or higher) chance for a case; the token records acting on it | Levers | none |
| drift | whether a refitted model's pattern has moved: keep, refit, abandon or no_check_yet |
Run it | none |
| evidence | outcomes of cases acted on against cases not acted on | Did acting work? | none |
| journey | one of five things a reader does: try it, ask your data, ship a product, run it, prove it | Introduction | none |