# Public-release checklist

## Completed in this sanitized copy

- [x] Preserved the original experiment directory unchanged.
- [x] Created a separate sibling copy for release preparation.
- [x] Removed logs, caches, `.nfs*` files, editor state, generated API responses, predictions, plots, baseline checkpoints, training outputs, previous experiment results, and raw/full dataset payloads.
- [x] Removed broken absolute symlinks and private machine-operation scripts.
- [x] Retained evaluation GT subsets and RQ3 simulation source without historical result outputs.
- [x] Added an honest repository overview, pipeline map, asset inventory, third-party notice, secret-safe example environment file, and release-focused `.gitignore`.
- [x] Centralized default/paper model profiles and switched the default path to OpenRouter.
- [x] Replaced source-edited APE JSONL slicing with CLI range/shard arguments.
- [x] Added offline unit tests for reflection crops, APE output shaping, provider config, and evaluation conversion.
- [x] Added configurable `run_vlm.py` and `run_ape.py` entrypoints with explicit data, output, model, APE, range, shard, resume, and reflection settings.
- [x] Added isolated shard filenames, zero-detection progress tracking, and a fail-closed shard merger with overlap/coverage validation.
- [x] Rebased legacy `method*.py` path construction on the current checkout while preserving historical stage branches and comments.
- [x] Added machine-readable manuscript author and ORCID metadata in `CITATION.cff`.
- [x] Bundled the APE-L\(_D\) checkpoint needed by the main detector and recorded its checksum in `docs/MODEL_MANIFEST.md`.
- [x] Configured Hugging Face/Git LFS tracking for the bundled checkpoint.
- [x] Added public asset verification helpers without committing private data, historical caches, generated predictions, or compiled APE artifacts.
- [x] Added environment/build notes without pretending the CUDA/APE stack is a universal lockfile.
- [x] Ran one-image live smoke tests for the default OpenRouter VLM, APE-L\(_D\), and one PII.5–PII.7 advisor round without retaining generated artifacts.
- [x] Added fail-closed offline/read-only semantic evaluation controls.
- [x] Removed the evaluator's shared temporary-file race with per-process paths and cleanup-on-failure tests.
- [x] Kept prior RQ3 raw/final experiment runs outside the release copy.
- [x] Added deterministic question-manifest generation from mounted screenshots.
- [x] Added optional pinned app metadata cache support for reproducible VLM prompts without live Steam drift.
- [x] Added project-local agent guidance forbidding direct OpenRouter key retrieval or exposure.
- [x] Added the published PACMSE/FSE DOI, article metadata, BibTeX, and preferred paper citation.
- [x] Added explicit remote-provider screenshot privacy guidance and fail-closed path/cache validation.
- [x] Added automatic SHA-256 verification before the released PyTorch checkpoint is loaded.
- [x] Added a deterministic one-image control-path smoke that uses no API or model weights and retains no outputs.

## Required before publication

- [ ] Decide and add the repository license.
- [ ] Resolve every `UNKNOWN` and license mismatch in `docs/THIRD_PARTY.md`; either pin and clear each retained vendor/asset or remove it from the chosen public scope.
- [ ] Add the final public repository/contact/demo/model links; add the dataset link when the dataset is published.
- [ ] Publish the dataset assets with versions, checksums, schemas, and licenses; preserve the bundled model manifest and attribution in the final Hugging Face revision.
- [ ] Pin a validated paper-scale CUDA/APE environment from a fresh public clone; resolve conflicting or duplicate dependency pins after the final target GPU host is selected.
- [ ] Add deterministic regression fixtures for the published paper tables without committing generated predictions or the 548 MB embedding cache.
- [ ] Decide how to repair or disclose the dynamic-coverage `sample` protocol; new comparisons must use one fixed GT image set and report coverage.
- [ ] Recover model/round provenance for the ablation artifacts and replace the non-model-controlled `w/o Context` comparison.
- [ ] Recover the three LMM rounds and verify the averaging procedure claimed in the paper.
- [ ] Reconcile the abstract's `103.1%` RQ3 coverage increase with the stored K=60 gains, or state the exact split and time point used for that percentage.
- [ ] Run a fresh-clone installation, inference smoke test, evaluation smoke test, and full selected-paper experiment.
- [ ] Run final secret, large-file, license, and personal-path scans on the exact commit to be published.
- [ ] Create a versioned tag/release with known limitations and asset revisions.