diff --git a/docs/review/datacite_dropped_by_remote_ls_2026-08-08.md b/docs/review/datacite_dropped_by_remote_ls_2026-08-08.md
index bee9c99..6cf57de 100644
--- a/docs/review/datacite_dropped_by_remote_ls_2026-08-08.md
+++ b/docs/review/datacite_dropped_by_remote_ls_2026-08-08.md
@@ -114,6 +114,64 @@ blanked: `descriptions` (preferring `descriptionType: Abstract`), `creators`,
 whole block alongside the derived columns, so an unusual or partial record
 costs us nothing.
 
+## Backwards compatibility: sidecars vary, and already carry the fields
+
+Two things we checked before asking, because both change the shape of the work.
+
+**1. The mirrored (file-based) path already has the data too.** We assumed this
+ask only helped the live LEXIS path and that mirrored datasets would need a
+re-pull. Not so — listing `owi-up/collectionName=main` gives 43 metadata keys
+per dataset, *more* than the LEXIS path's 42, including `creators`,
+`descriptions`, `publisher`, `publicationYear`, `rightsList` and
+`relatedIdentifiers`. **So no sidecar migration is required for this request.**
+The fields are in the sidecars that were written at pull time; only the HTTP
+projection loses them.
+
+**2. Sidecar format is not uniform, and a fix must not assume it is.** Across
+the 585 mirrored `main` datasets on this one deployment there are **six
+distinct key sets**, from 33 to 43 keys:
+
+| datasets | keys | difference from the most common shape |
+| --- | --- | --- |
+| 192 | 41 | — (baseline) |
+| 189 | 43 | `+workflow_ex_id`, `+workflow_id` |
+| 168 | 40 | `+provenance`, `−fundingReferences`, `−relatedIdentifiers` |
+| 24 | 42 | `+provenance`, `+workflow_ex_id`, `+workflow_id`, `−fundingReferences`, `−relatedIdentifiers` |
+| 11 | 41 | `+workflow_ex_id`, `+workflow_id`, `−fundingReferences`, `−relatedIdentifiers` |
+| **1** | **33** | also missing `descriptions`, `publisher`, `publicationYear`, `rightsList`, `schema`, `types`, `contributors`, `dates` |
+
+Per-field presence over the same 585:
+
+```
+creators            585/585      schema              584/585
+descriptions        584/585      publisher           584/585
+publicationYear     584/585      rightsList          584/585
+relatedIdentifiers  381/585      <- only 65%
+```
+
+`schema` is `http://datacite.org/schema/kernel-4` on 584 and absent on the
+remaining one.
+
+What that implies for the implementation:
+
+- **Every field must be optional.** `relatedIdentifiers` is missing from a third
+  of them and one sidecar has almost no provenance at all. A widened record must
+  omit what is absent rather than emit `null`-filled or partially-constructed
+  objects — a consumer has to be able to tell "not published" from "not
+  carried".
+- **Do not key on a fixed key set or a key count.** Six shapes exist in *one*
+  mirror; other installations will have others.
+- **`schema` is the right version marker** where present, and its absence must
+  be tolerated rather than treated as an error.
+- **Do not require a re-pull or a sidecar rewrite.** Older mirrors are not ours
+  to migrate, and a change that only works for freshly pulled data leaves every
+  existing mirror behind.
+
+Sidecar migration *is* available to us if you would rather normalise the format
+— `owilix remote reindex` already rebuilds the index from sidecars, so there is
+a natural place for it — but please do not make it a **precondition** for
+reading the fields, for the reason above.
+
 ## While you are in that code: `DatasetMetadata` is not a usable mapping
 
 Separate, small, and it will bite whoever implements the above — because the