Not planned

These ideas stay off the library surface. They are listed so the numbered backlog is not read as “everything in the GP literature.”

See the overview for what is planned, and Under consideration for ideas that may still join that table.

Idea	Why not
PushGP, Cartesian GP, or linear GP as a second public genome	Prefix trees plus the tape already linearize a program. A second representation needs a caller who cannot use trees or columnar arrays. A private Push policy that only sees summaries is a different track: Push GP.
Transformer or LLM mutation	Heavy optional dependencies, unstable operators, and they do not compose with the tape. A recipe in a notebook is enough.
A catalog of named domain indicators as primitives	Composable windows and user opcodes. Named catalogs age badly.
Built-in domain fitness functions	They need application state the library does not own.
Learned quality-diversity / meta-BBO	A research paper, not a toolbox function. Item 22 is a semantic descriptor and nearest-neighbor, not a trained QD model.
Interactive / human-in-the-loop evolution	A notebook over archive + tree_to_infix. Machine-checkable pressure is items 22–23 and 26, not clicks.
Switching persistence off dill, or replacing creator with dataclasses	Forbidden by the project contract.
An in-tree C / Cython rewrite of operators, CMA, or selection	Not the bottleneck; those modules must stay readable. Revisit an optional tape backend only after items 3, 4, 9, and 32 are in and a profile still points at interpret.