diff --git a/docs/adr/032-person-delete-db-level-integrity.md b/docs/adr/032-person-delete-db-level-integrity.md
new file mode 100644
index 00000000..de44a50c
--- /dev/null
+++ b/docs/adr/032-person-delete-db-level-integrity.md
@@ -0,0 +1,62 @@
+# ADR-032 — Person-delete referential integrity lives in the database, and the cascade never reaches `documents`
+
+**Date:** 2026-06-06
+**Status:** Accepted
+**Issue:** #684 (move person-delete FK detach to database-level `ON DELETE`)
+**Milestone:** —
+
+---
+
+## Context
+
+Deleting a `Person` had to detach the two FKs into `persons` that lacked any `ON DELETE`
+behaviour: `documents.sender_id` and `document_receivers.person_id` (both from V1).
+`PersonService.deletePerson` and `mergePersons` did this in Java — nulling the sender and
+deleting receiver join rows before `deleteById` — so the integrity guarantee lived in
+application code. Any other delete path (a future endpoint, a manual `psql`, a batch job)
+could still orphan rows or fail with an FK-violation 500.
+
+A related soft reference made it worse: `transcription_block_mentioned_persons.person_id`
+was a UUID column with **no FK** (V56, a deliberate "no FK" choice), so a person delete left
+dangling `@`-mention rows. The literal `@DisplayName` lives in `transcription_blocks.text`,
+so only the *link* was ever at stake — not the visible name.
+
+## Decision
+
+Move person-delete integrity into the database (migration V71) and thin the service to a
+plain `deleteById`:
+
+- `documents.sender_id` → `ON DELETE SET NULL` (`documents.senderText` preserves the raw
+  textual attribution, so nulling the link loses no historical record).
+- `document_receivers.person_id` → `ON DELETE CASCADE` (the symmetric completion of V14,
+  which gave the `document_id` side the same).
+- `transcription_block_mentioned_persons.person_id` → a real FK with `ON DELETE CASCADE`,
+  reversing V56's "no FK" decision. The read renderer already degrades a `@DisplayName` with
+  no sidecar row to plain escaped text, so removing the link is invisible to the reader.
+
+**Cascade-boundary invariant:** the cascade stays strictly at the join/reference layer and
+**never reaches `documents` rows** — a cascade into `documents` would destroy historical
+letters. This is pinned by a non-negotiable document-survival assertion in
+`PersonRepositoryTest`.
+
+## Consequences
+
+- A person delete is safe from every path, not just `PersonService`. The service and merge
+  stay thin (`deleteById` + the cascade); `reassignSenderToNull` and `deleteReceiverReferences`
+  are deleted.
+- This *fixes* the pre-existing dead-link-on-deleted-person case — it is not a purely
+  invisible refactor.
+- DB cascades run below `AuditService`, so the row-level cleanup is intentionally not
+  audit-logged; the person-delete action itself is still logged at the service layer.
+- The V71 FK validation requires cleaning pre-existing orphan mention rows first; the
+  migration does this in a `DO` block that logs the purge count via `RAISE NOTICE`.
+
+## Alternatives considered
+
+- **Keep integrity in Java** — rejected; it only protects the one code path and re-breaks the
+  moment a second delete path appears.
+- **Cascade `documents.sender_id`** — rejected; it would delete historical letters when a
+  sender is removed. `SET NULL` keeps the letter and its `senderText`.
+- **Leave the mention sidecar FK-less (honour V56)** — rejected; the "no FK" rationale was
+  stale, the name survives in the block text regardless, and the FK removes the orphan-row
+  class of bug.