Introduction

Fang is a Python-embedded language for circuit boards and the kernel that holds what you write. You author intent: typed interfaces, parameters carrying units and constraints. The kernel elaborates that into one canonical model, refuses anything it cannot justify and lowers what survives into netlists, views, simulation decks and KiCad.

It is the layer under copperhead. Copperhead is the agent you talk to. Fang is what it writes and what checks the writing.

The shape of it

class SensorBoard(System):
    regulator = Regulator(output_voltage=3.3 * V, package="SOT-23-5")
    mcu = MCU(package="LQFP-64")
    imu = IMU(package="LGA-14")

    def architecture(self):
        self.mcu.i2c >> self.imu.i2c

    def constraints(self):
        require(self.regulator.output_voltage == 3.3 * V)

self.mcu.i2c >> self.imu.i2c is an interface connection, not a pin assignment. The kernel lowers it onto PB8 → SCL and PB9 → SDA, the same way every time. The MCU offered PB6 and PB7 as alternatives, so the choice becomes a decision entity naming what it rejected.

Pin assignment is a lowering result. It is never an authoring input, which is what makes late assignment and part substitution tractable.

What it will not do

Most of the value is in the answers the kernel declines to give.

• An unknown datasheet number makes a check undecided. That is a third truth value, never a quiet pass. Whether undecided blocks is the commit gate’s policy decision, not the evaluator’s.
• A dimensionally invalid expression is rejected where it is written, so it can never be stored, let alone evaluated.
• A missing simulator reports unsupported rather than substituting a model.
• A CAD construct the adapter does not model is named in the import report rather than dropped.
• A generic part carries no manufacturer until one is selected.

Nothing is invented to make an answer available.