dataset.py

Sentinel/Lab/sentinel_lab/dataset.py


Family	Lab (Python)
Version	—
Size	297 lines
Scope	public — ships in `sentinel-suite`
Documented by	no doc tracks this artifact

Rendered from the published copy in sentinel-suite/src/, not the author’s private tree — so this page describes the file you actually have.

What the file says about itself

Load Sentinel excursion JSONL into a training frame.

Schema-tolerant: reads 1.2 (no decision vector) and 1.3 (with it). Rows that
predate 1.3 simply carry NaN in the voter columns and are dropped by the weight
trainer -- but they remain fully usable for the calibration curve, which needs
nothing but `conviction`.

The one modelling decision that lives here: FOLD BY DIRECTION.

    x_i = vote_i * dir

A long verdict with EYE=+1 and a short verdict with EYE=-1 are the SAME evidence
("the Eye agreed with the taken side"). Folding halves the feature space, doubles
effective N, and -- because `dir == sign(netScore)` for any verdict -- makes the
fitted coefficient directly comparable to the Council's hand-set `WeightEye`.
It drops straight into Model.conf.