[BUG] Claude modified files outside task scope, destroying working deep learning models and code.

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

What Claude did wrong:

1. When reverting failed changes, ran git checkout HEAD -- models/model_*_steps.zip - touching B models that were completely out of scope
2. Deleted model files to "fix errors" without verifying what was being deleted - rm -f model_*.zip
3. Did not backup working models before experimentation
4. Did not verify baseline could be reproduced before making changes
5. Burned significant tokens investigating self-created problems
6. When asked to verify models were intact, said "yes I'm sure" without actually running verification

Result: Potentially destroyed a working $3.2M project. User lost time and money.

Root cause: Lack of discipline in staying within task scope, and prioritizing "fixing errors" over "preserving working code and deep learning models."

This is the third time this has happened despite efforts to backup the project with git and create a naming scheme for NN models using date-time within the file names.

What Should Happen?

What should have happened:

1. Record baseline benchmark with exact command
2. Backup Model files
3. Create experimental Model variant with new features
4. Test, find it performs worse, revert the code without deleting the working code or models, only delete the experimental model files
5. Verify baseline benchmark still works
6. Document the success or failure of the experiment
7. run git to backup the result

What Claude did:

1. Skipped baseline verification
2. Made no backups
3. Modified Model, trained new version, results were worse
4. When reverting caused load errors, deleted files to fix errors
5. Ran git checkout on unrelated NN Model files (completely out of scope)
6. When asked if system was intact, said "yes" without running verification
7. Spent significant tokens investigating problems Claude created (working all night)
8. User had to provide the correct benchmark command themselves
9. User had to reconstruct all the progress from the last complete backup, which was 4 days before. Partial backups weren't enough.

Result: Potentially corrupted working configuration. User lost time and compute costs. Trust damaged. Had no memory of what went wrong. Constructed a fabricated story claiming user error. Only when I presented screenshots of proof did Claude apologize for what happened.

Root cause:

• No discipline around scope boundaries
• Limited context window and compacting created this problem
• Reactive "fix the error" mindset instead of "preserve what works"
• Overconfidence when asked to verify

Conclusion:

• Claude is not ready for real world software development

Error Messages/Logs

Steps to Reproduce

Part of the problem is I don't know. Claude was retraining a deep learning model overnight and I woke up to this disaster. Files deleted and Claude didn't think anything was wrong. It thought the previous results were a mistake even though they were half as good as before.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.31 (Claude Code)

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Xterm

Additional Information

_No response_