Which build string configurations in environment.yml lock binary reproducibility across platforms?

Confirmed facts

A build string pin (e.g., package=1.2.3=h1234567_0) selects one exact package build from a specific channel for the platform on which the solve runs. This is the strongest reproducibility mechanism that environment.yml can express.

Likely explanation

Because build hashes are platform‑specific, the same string rarely exists for linux‑64, osx‑64, osx‑arm64 and win‑64. Pinning a build string in a shared environment.yml therefore locks the binary only for the OS where that hash is available; on other platforms conda will either fail to find the package or fall back to a different build, breaking binary reproducibility.

Steps to achieve cross‑platform locking

1. Fix the channel list and set channel_priority: strict (or equivalent) in environment.yml or .condarc so that identical version+build names always come from the same channel.
2. Pin exact build strings for the packages you need locked (e.g., python=3.11.4=h1234567_0). Remember these pins are valid only for the platform you are solving on.
3. Generate a platform‑specific lock file using conda-lock or run conda env export --no-builds on each target OS and commit those lock files to version control.
4. When creating the environment, install from the lock file for the current OS (e.g., conda-lock install --conda conda-lock.yml).
5. Verify reproducibility by running conda list --explicit on two machines of the same OS and comparing the output; the lists should match exactly.

What to check if you cannot use lock files

If you are unable to generate per‑platform lock files, the only way to approximate reproducibility with a shared environment.yml is to:

• Pin only the version (no build string) and rely on a stable, immutable channel (e.g., a mirrored conda‑forge snapshot).
• Accept that minor binary differences may appear across OSes and test accordingly.

To decide which approach fits your workflow, please confirm: Are you able to generate per‑platform lock files using conda-lock or conda env export for each target operating system? If the answer is no, the recommendation shifts to version‑only pinning with an immutable channel and acceptance of possible binary variance.