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.

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.