Citations, not dependencies
The map does not run these codes and does not contain them: it cites them. Every representation rail on the map records which code serves a quantity, through which interface, with a citation, a DOI where verifiable, and the code's license. This page is that bibliography, rendered from the same data file the map reads (docs/data/codes.json).
Most codes on this page compute numbers; Lean 4 checks reasoning. The map's
dimensional structure and operator identities are stated as
Lean theorems and compiled against
Mathlib and physlib, so a slice of the map is not just tested but proven
(zero sorrys is the proof). Read
the verified layer for the tiers and the honest
limits: a proof here checks dimensional consistency or an algebraic
identity, not that a formula is the right physics or that any value is
accurate.
The three works this rests on:
One entry per code, straight from codes.json. The count after
each name is how many quantities that code serves on the map; every entry
links to the map filtered to that code's rails. Entries sort by rail count,
then by name.
Loading the bibliography from ../data/codes.json…
A rail records interface facts (keywords, file formats, units), citations, and published equations, written in our own words. The map never vendors or copies a code's source, so no code's license flows into this repository: the kernel is Apache 2.0, the map data is CC BY 4.0. Copyleft codes, GPUMD under GPL-3.0 among them, are cited and described exactly like permissively licensed ones. Describing an interface for interoperability and citing a paper are not acts the GPL restricts.
Running any code stays under that code's own license, on the runner's side.
Every rail's recorded license is enforced in CI:
tests/test_code_credits.py reads
omai/representation/credits.py, licenses come from the codes'
own LICENSE files, and DOIs are never invented.