A function, constructed by context.

The candidate rules never train or change. Your examples only determine which rule can be used.

Fixed candidate rule spacePROBE x = 4

DOUBLE 2x SQUARE x² CONSTANT 4

Load contextObserve input-output examples.

PredictChoose a rule or underdetermined.

FilterKeep every rule matching all examples.

ProbeCompare the remaining outputs at x = 4.

Choose the evidence before you choose the function.

One example can fit several rules. More context can collapse that ambiguity without updating any rule.

Context examples

Your prediction

Context loaded: (1 → 2), (2 → 4), (3 → 6)
Predict before revealing the compatible set.

The forge is waiting for a prediction.

Context does selection work only after you commit to an expectation.

2xx²4

This is an illustrative rule library, not a neural model or evidence about a specific circuit. It isolates one idea: context can select or construct usable behavior without gradient updates during the interaction.