# Workspace cleanup — 2026-08-17

This manifest records the conservative post-competition cleanup policy.

## Preserved

- All training/evaluation datasets, including `experiment_4`, `external_eval`,
  `rasam-dataset-main`, `_mh`, PAGE-XML/image corpora, and `tmp/exp8`.
- All Git worktrees and their branch history.
- Deployed recognition, segmentation, specialist, and language models.
- The local database and `.env` needed to reproduce the running application;
  both remain ignored by Git.
- Source-retrieval texts, reports, benchmark summaries, and the prepared
  Hugging Face package.
- The shared backend virtual environment used by the demo launcher.

## Removed as regenerable or superseded

- Repeated `_wheels` download caches in the four worktrees.
- The isolated EasyOCR pilot environment and an unrelated temporary Python
  environment.
- Playwright, Python bytecode, and pytest caches; browser screenshots/output.
- Untracked competition ZIPs and duplicate unpacked presentation bundles.
- Temporary server logs, OCR book probes, RAC/top-k caches, and abandoned
  server-patch build directories.
- The pre-competition copy and the pre-segmentation source backup after
  confirming that the active models and Git history are present.
- Root-level scratch JSON/TXT/log files and stale PID files.

No training dataset, deployed model, language model, source corpus, Git
history, or runtime database was intentionally removed.