Words use the plate.
Learning moves it.

Hold inference fixed, then run one real gradient step. See exactly why changing a target can reverse the parameter update while the forward prediction stays identical.

SIGMOID + BINARY CROSS-ENTROPYSCALAR MECHANISM · NOT LORA
LOGIT · wx+b—Current parameter evidence
FORWARD PREDICTION—Same before any learning
LOSS · BCE—Target-relative penalty
ERROR SIGNAL · p-y—Backward signal

Prompt-only inference

prediction = —
weight before = —
weight after = —
optimizer invoked = false
PARAMETER STATEUNCHANGED

Weight update + counterfactual

dW = —
new W = —
flipped-label dW = —
flipped new W = —
UPDATE DIRECTIONS—

Boundary. Ordinary inference does not run an optimizer over parameters. This scalar trace explains that mechanism only; it is not LoRA, a language model, a values benchmark, Tinker integration, or evidence about model ownership.

READY FOR AUDIT

One input. Same prediction. Opposite learning directions.

0.622459PROMPT-ONLY · WEIGHTS UNCHANGED
w ↑LABEL 1 UPDATE
w ↓LABEL 0 UPDATE

The forward prediction does not know the target. The labeled loss does. TensorFlow.js carries that target-relative error backward, and gradient descent moves the parameter in the resulting direction.

Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.