Research Codebase

Build the harness around the science: a fixed pipeline contract, a runner that executes many configurations, and an aggregator that turns finished runs into the paper's table. The user supplies the mathematics; this skill makes sure the mathematics gets run many times and reported correctly.

Two rules dominate:

• Fix the contract up front, then never edit shared code again. Adding an experiment must be one line in a list or one entry in a dict. If it requires editing the entry point, the contract is wrong.
• Provenance is a mechanism, not a discipline. Every run writes what produced it. Nothing depends on the user remembering to record anything.

The second rule is why "add structure only when the pain justifies it" is not the stance here. Deferring structure assumes someone who will recognize the pain and can perform the refactor. When the user cannot code, the refactor never happens — the project just accumulates train_v2_final_REAL.py. Give the small fixed skeleton once, correct, and forbid growing it.