---

name: customization-and-constraints description: "Customize PySR search space, custom operators, losses, constraints, units, and mutation or plugin behavior." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

customization-and-constraints

Use this sub-skill when the user wants to change how PySR searches, scores, or rejects expressions rather than simply fitting a standard model.

Route here for

• custom unary or binary operators, or arity 3+ operators via operators
• Float32/Float64-safe Julia operator definitions
• typed NaN guards for invalid operator domains
• extra_sympy_mappings for custom operators
• elementwise_loss, loss_function, or loss_function_expression
• operator constraints, nested_constraints, or complexity shaping
• dimensional constraints, X_units, y_units, or dimensionless_constants_only
• mutation and plugin configuration objects

Route elsewhere for

• ordinary fitting, Pareto selection, or search debugging -> fit-and-diagnose
• known algebraic skeletons, templates, or structured expressions -> structured-expressions
• export, reload, JAX, PyTorch, SymPy, or LaTeX artifacts -> export-and-artifacts
• startup, threading, backend, or cluster issues -> runtime-and-scaling

Working rules

1. Match operator arity to the constructor path before drafting code.
2. Require a SymPy mapping for every custom operator that must export cleanly.
3. Guard invalid domains with typed NaN, not Python-style exceptions.
4. Keep custom losses type-stable; use loss_scale="linear" when negative losses are possible.
5. Prefer finite structural penalties over all-or-nothing Inf whenever evolution needs a gradient toward compliance.
6. Leave slack in maxsize, operator constraints, and units so intermediate expressions remain reachable.
7. Use mutations / default_mutations and plugins / default_plugins when the user wants runtime tuning objects instead of legacy weight knobs.

Start here

• references/operators-and-losses.md for custom operators and loss modes.
• references/constraints-and-units.md for constraints, complexity shaping, and units.
• references/api-reference.md for current constructor and fit signatures.
• references/troubleshooting.md for common failures and fixes.
• scripts/validate_custom_operator.py for a safe static checklist before implementation.

Expected outcome

Return a concrete customization plan or validated operator/loss/constraint draft, and state any unresolved domain, export, or unit caveats explicitly.