Primitives

Each AI column asks one kind of question. All the questions for a row see the same row state and are answered independently. Jev answers with numbers and ids, never prose, and AI Fill shows only what Jev returned and what you configured.

Choice

A Choice picks one of your options.

persona: {
    primitive: "choice",
    instructions: "Which buyer persona best describes this contact?",
    sources: ["title", "notes"],
    options: {
        champion: { description: "Drives the purchase internally", label: "Champion" },
        economic: { description: "Controls the budget and signs off on spend", label: "Economic buyer" },
        none: { description: "None of these fit", label: "None of the above", outcome: "none" },
    },
}

The answer:

{
    "type": "choice",
    "choice": "economic",
    "probabilities": { "champion": 0.06, "economic": 0.91, "none": 0.03 },
    "confidence": 0.84
}

• probabilities has every option, and they sum to 1. choice is always the most probable option.
• The probability of the selected option is how much of the answer's weight is on it.
• The confidence is a separate number the model reports about its answer as a whole. It is higher when the probability is concentrated on one option. It isn't a probability.
• Option ids and descriptions are what Jev sees. label is what users see, and value is what is written (default label ?? id).
• An option with outcome: "none" or "unknown" is a semantic outcome, such as "none of these" or "not enough information". It is a real answer, not an error and not a withheld result. It is drawn in muted italics with a ◇ marker, and accepting it writes nothing unless the option sets a value.

Score

A Score places the row on your rubric.

seniority: {
    primitive: "score",
    instructions: "How senior is this contact?",
    sources: ["title"],
    levels: ["Individual contributor", "Manager", "Director", "Vice president or C-level"],
    output: { store: "level-label" },
}

The answer:

{
    "type": "score",
    "score": 2.88,
    "legend": { "0": "Individual contributor", "1": "Manager", "2": "Director", "3": "Vice president or C-level" },
    "probabilities": { "0": 0, "1": 0.02, "2": 0.08, "3": 0.9 },
    "confidence": 0.88
}

• The score is a position on the rubric, from 0 to levels − 1. It is the probability-weighted average of the level indexes. It is not a percentage, and it isn't a probability.
• A high score says where the weight sits, not how sure the model is. When the probability is split between distant levels, the score can land between them, on a level that almost no probability supports. Read the confidence and the level probabilities too.
• output.store decides what is written: the score, the level index, the level's label or value, or your own function. output.levelFrom decides which level: the one nearest the score (half up, the default) or the most probable one. With probabilities [0.45, 0.05, 0.05, 0.45] the score is 1.5, which is nearest to level 2, while the most probable level is 0 (ties go to the lower level).

Noul

A Noul answers a yes/no question with one number.

ownsBudget: {
    primitive: "noul",
    instructions: "Does this contact own a budget?",
    criteria: { true: "Signs off on spend", false: "No spending authority" },
    sources: ["title", "notes"],
    output: { store: "boolean", bands: { falseAtOrBelow: 0.2, trueAtOrAbove: 0.8, between: "review" } },
}

The answer is { "type": "noul", "noul": 0.73 }: the probability that the answer is yes.

• There is no confidence. A policy can't gate a Noul on confidence, and the inspector doesn't show one.
• A value near 0 is a strong no, not a failure. 0.02 is a usable "No".
• To write yes or no, set bands. At or below falseAtOrBelow is no, at or above trueAtOrAbove is yes, and in between is "Uncertain": it goes to review (or is withheld with between: "withhold"), and it is never shown as "No". Accepting an uncertain result writes nothing.
• Without bands, store: "probability" writes the probability itself.

Probability is not confidence

For a Choice they are two different numbers, and they can disagree:

A policy gate reads only the measure it names: minProbability never looks at confidence, and minConfidence never looks at probability. So with show: { minProbability: 0.8 } the first row is shown and the second is withheld. See Result policies.

Neither number is measured accuracy. They are what the model reports. Evaluate thresholds on your own data before you rely on them.

Answer, display and value

AI Fill keeps three things apart for every result:

• the answer: Jev's raw answer, validated, with the model id that produced it;
• the display text: what the cell and the inspector show, for example Economic buyer · 0.91;
• the value: what is written when the result is accepted, for example "Economic buyer", or nothing for a semantic outcome or an uncertain Noul.

output.format changes the display, and output.toCell changes how the value becomes a cell. Neither changes the answer, and changing them re-decides stored answers without a new request.

Validation of answers

Every answer is checked before any policy runs. An answer is malformed when it has the wrong type, a choice that isn't an option or isn't the most probable one, probabilities that don't match the options or levels or don't sum to 1 ± 0.01, a value outside [0, 1] or a Score outside [0, levels − 1], or no model id. A malformed answer is an error, and nothing is written.