Examples define the target.
The original four labels encode an OR rule: only zero plus zero is QUIET. Training reduces error against those supplied targets.
Teach a network by changing the evidence.
Predict one decision. Train on four labeled examples. Then change exactly one label and watch the same architecture learn a different boundary.
LEARNED DECISION FIELDUNTRAINED
PROBE .9 / .1
0 / 0QUIET
0 / 1SIGNAL
1 / 0SIGNAL
1 / 1SIGNAL
YOUR PREDICTIONPROBE .9 / .1
What will the original labels teach?
INPUT [0.90, 0.10]
Loading the local numeric engine.
PROBE DECISION0 / 4 FIT
WAITING
CONFIDENCE --
H1 --H2 --H3 --H4 --
Make a prediction, then let the network update its weights from the labels.
NO TRAINED BOUNDARY YET
The label is not the law.
This tiny network exposes the full causal chain instead of hiding learning behind a score.
Weights create a boundary.
TensorFlow.js updates the same eight input weights, four hidden biases, four output weights, and one output bias until the labeled points separate.
A changed label changes the lesson.
Flip the east example and the target pattern becomes dependent on the second coordinate. The probe stays fixed; the learned evidence changes around it.
Carry the idea elsewhere.
A surprising model decision is a reason to inspect training evidence, not proof that the model found a hidden truth.