8.9 The DLINCS compiler

The computational architecture can be organized as a compiler with a small interface.

Compiling a deep learning sketch

1. Parse typed objects, generators, serial wiring, and tensor wiring.

2. Construct the quotient relations and designated universal properties.

3. Instantiate every generator as a fixed or parametrized module.

4. Compile base obstruction observers for the declared relations.

5. Lift admitted objects and arrows to tangent or Weil probes.

6. Compile reverse-mode updates and requests compositionally.

7. Quotient parameter gauges and presentation-null directions.

8. If an intrinsic metric is declared, solve the typed tangent repair on the quotient and record any curvature approximation.

9. Retract the tangent proposal to parameter space without conflating it with admission.

10. Return separate base, tangent, update, and request certificates.

This compiler need not force every module to be trainable. Fixed transitions, incidence maps, restrictions, and aggregation operators are parametrized maps with terminal parameter objects. Trainable modules and fixed structure can therefore coexist in one sketch.