Preconditioner/loss comparison for removing unwanted signals

[davidoj]

Add an experiment that compares various methods for attributing "semantic" information but not "style" information.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ luciaquirke
❌ David Johnston

---

David Johnston seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[claude]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.

[LouisYRYJ]

Hmmh I am wondering if from now on/in future we want to have experiments using bergson to be in their own repo? Otherwise we might get a lot of overhead?

[davidoj]

@luciaquirke wanted this in as an example, because she'd received the feedback that the repo had too few examples. Though I agree that we don't want to add every experiment someone does to the repo.

[luciaquirke]

This is fabulous, thank you so much! I updated the docs file to have more of a quickstart vibe. I'm pretty new to skills files so idk if I can review them effectively but it would be nice to put the ones specific to this experiment into a subdirectory because I can imagine it getting pretty hard to navigate quickly. Otherwise LGTM

My claude is cleaning up some format issues

[luciaquirke]

@claude could you check if the pre-commit succeeds and fix it if it fails, thanks

[claude]

Claude finished @luciaquirke's task in 4m 57s —— View job

---

Pre-commit check complete

• Check CI build status
• Verify no obvious formatting issues in changed files
• Monitor build for pre-commit failures

Summary

I checked the pre-commit status for this PR. Here's what I found:

Pre-commit checks: The repository uses pre-commit.ci (an external service) which runs automatically on pull requests. The GitHub Actions "build" workflow doesn't include pre-commit checks directly - it only runs pytest, pyright type checking, and builds the package.

Code formatting: I verified the changed files and found no obvious formatting issues:

• ✅ No trailing whitespace
• ✅ All files end with proper newlines
• ✅ No whitespace errors detected in the diff

Current status: The GitHub Actions build is still running (pytest and pyright). However, since pre-commit.ci is a separate service that runs independently, you should see its status in the PR checks section on GitHub once it completes.

Note: I wasn't able to run pre-commit run --all-files locally as it requires approval for command execution. However, based on my inspection of the code changes, the files appear to be properly formatted and should pass the pre-commit hooks (black, ruff, trailing-whitespace, end-of-file-fixer).

If pre-commit.ci does report any issues, it typically creates an automatic commit to fix them, or you can check the pre-commit.ci bot's comment for details.

---

| Build status