Add EmbodiedMemoryBench dataset card

README.md

@@ -0,0 +1,49 @@
+# EmbodiedMemorizerBench Full-2554
+
+This directory is the organized, canonical view of the original
+EmbodiedMemorizerBench data used by the full 2,554-episode evaluation.
+
+It intentionally excludes all sampled and supplementary suites, including
+Diagnostic-160, Confirm-400, Hard-interaction-120, OOD, counterfactual,
+stress, anti-scaffold, smoke, debug, and augmentation outputs.
+
+## Layout
+
+```text
+episodes/
+  passive_observation/        1,036 episodes
+  dynamic_tracking/           1,052 episodes
+  interaction_experience/       263 episodes
+  experience_generalization/    203 episodes
+manifests/
+  full2554.jsonl              canonical frozen membership
+  passive_observation.jsonl
+  dynamic_tracking.jsonl
+  interaction_experience.jsonl
+  experience_generalization.jsonl
+  audit_report.json
+browse.tsv                    one searchable row per episode
+metadata.json                 view provenance and counts
+verification_report.json      link and episode-hash verification
+build_view.py                 idempotent builder and verifier
+```
+
+Each entry below `episodes/` is a relative symbolic link to an immutable
+canonical episode directory. Opening a link gives direct access to the
+episode JSON and all associated context images. The source trajectories are
+not moved, renamed, copied, or modified, so existing evaluations continue to
+use their original paths.
+
+The only authority for dataset membership is `manifests/full2554.jsonl`.
+Do not add arbitrary directories from `output/trajectories`, which also
+contains non-canonical development artifacts.
+
+Run the integrity check from this directory with:
+
+```bash
+python3 build_view.py --verify-only
+```
+
+If a physical standalone copy is needed for transfer or release, materialize
+the links into a new destination rather than changing this view or the source
+data. For example, `rsync -aL` follows the links and copies their contents.