diff --git a/internal/adapters/store/migrations/019_uploaded_transcripts_reconciliation.down.sql b/internal/adapters/store/migrations/019_uploaded_transcripts_reconciliation.down.sql
new file mode 100644
index 0000000..46eba75
--- /dev/null
+++ b/internal/adapters/store/migrations/019_uploaded_transcripts_reconciliation.down.sql
@@ -0,0 +1 @@
+ALTER TABLE uploaded_transcripts DROP COLUMN reconciliation;
diff --git a/internal/adapters/store/migrations/019_uploaded_transcripts_reconciliation.up.sql b/internal/adapters/store/migrations/019_uploaded_transcripts_reconciliation.up.sql
new file mode 100644
index 0000000..433870a
--- /dev/null
+++ b/internal/adapters/store/migrations/019_uploaded_transcripts_reconciliation.up.sql
@@ -0,0 +1,12 @@
+-- Migration 019: how a merged transcript came about (ADR-034).
+--
+-- When two engines transcribed the same audio and a local model chose between their
+-- disagreements, this records the status, the exact counts, and a bounded list of the
+-- contested spots (changed or unsure, each with a position and both readings) so the
+-- page can say what was done and the user can check the uncertain spots against the
+-- audio. NULL means a single engine: no second opinion, and the page must not claim one.
+--
+-- It lives on the already-FORCE'd, RLS-scoped uploaded_transcripts, so the exact-set
+-- isolation test is unaffected. The contested spots hold fragments of the private
+-- transcript, which is why they are here and nowhere else.
+ALTER TABLE uploaded_transcripts ADD COLUMN reconciliation JSONB;