Add mirror README with takedown policy

README.md

@@ -0,0 +1,51 @@
+---
+license: other
+license_name: mixed-upstream
+license_link: https://github.com/deepsafehq/deepsafe-bench/blob/main/THIRD_PARTY_NOTICES.md
+tags:
+  - deepfake-detection
+  - model-zoo
+  - mirror
+---
+
+# DeepSafe Model Code Mirror
+
+Inference code for the detection models used by
+[DeepSafe](https://github.com/deepsafehq/deepsafe-bench).
+
+**This is a mirror. None of this code is ours.** Every model is the work of its
+original authors and retains its own license. See each model's `SOURCE.md`.
+
+## Why this exists
+
+Research code and paper weight links rot. DeepSafe lost one model entirely when
+its weights were deleted from the host the paper linked to. Mirroring means
+`setup.sh` still works in three years.
+
+## Layout
+
+```
+clean/<modality>/<model>/     inference code only, ~16 MB total
+  SOURCE.md                   upstream URL, mirror date, license
+  LICENSE                     original, verbatim, where upstream ships one
+archive/<modality>__<model>.tar.gz   full untouched snapshot, ~218 MB total
+```
+
+`setup.sh` pulls `clean/**` only. The archive exists so nothing is lost: if you
+need training scripts, configs, or paper figures that inference does not, they
+are there.
+
+Stripped from `clean/`: README media, demo videos, training splits, dataset
+protocol lists, upstream CI config and tests, and a vendored copy of fairseq
+that nothing referenced. Verified by AST analysis that no intra-package import
+broke.
+
+## Takedown
+
+If you authored any of this and want it gone, open an issue at
+https://github.com/deepsafehq/deepsafe-bench and it is removed within 48 hours,
+no questions asked and no justification needed. We mirror to preserve
+reproducibility, not to claim ownership.
+
+If you would rather add a license than have it removed, we much prefer that
+outcome and will update our notices the same day.