Boolean

Boolean handles bool | None values without building or synchronizing a vocabulary. It works at any nesting depth represented by the surrounding Branch nodes.

import json2vec as jv

model = jv.Model(
    d_model=16,
    n_layers=1,
    n_heads=4,
    events=jv.Branch(length=8, succeeded=jv.Boolean),
    converted=jv.Boolean(target=True),
)

Representation

Visible content uses a fixed scalar representation:

Source value	Content
False	-1
True	1
None	0

The normal tensorfield state channel distinguishes null, padded, and masked values, all of which have zero content. A fixed three-row embedding buffer maps the content into the model width. Every dimension of its False, zero, and True rows is -1, 0, and 1, respectively. The buffer follows the model across devices and checkpoints but is not optimized. There is no vocabulary, top-k search, or vocabulary synchronization callback.

State and Boolean-class counters track observed training frequencies. Their frequency-derived weights balance state cross entropy and Boolean content loss.

Targets use binary cross entropy with logits. Only valued targets contribute to the content loss and binary metrics; null, padded, and masked targets contribute only to state prediction. Training logs epoch-level content accuracy, precision, recall, and AUC using stateful TorchMetrics, in addition to state accuracy and the state/content losses.

threshold defaults to 0.5 and controls the decision boundary used by accuracy, precision, recall, and the Boolean value in prediction output. AUC continues to use unthresholded probabilities.

Predictions contain the thresholded Boolean value and probability of True:

{
    "record/converted": {
        "state": {"valued": ..., "null": ..., "padded": ..., "masked": ..., "other": ...},
        "content": {"value": ..., "probability": ...},
    }
}