diff --git a/internal/adapters/store/migrations/020_uploaded_transcripts_progress.down.sql b/internal/adapters/store/migrations/020_uploaded_transcripts_progress.down.sql
new file mode 100644
index 0000000..f80bcd2
--- /dev/null
+++ b/internal/adapters/store/migrations/020_uploaded_transcripts_progress.down.sql
@@ -0,0 +1,4 @@
+ALTER TABLE uploaded_transcripts
+    DROP COLUMN alive_at,
+    DROP COLUMN progress_at,
+    DROP COLUMN progress;
diff --git a/internal/adapters/store/migrations/020_uploaded_transcripts_progress.up.sql b/internal/adapters/store/migrations/020_uploaded_transcripts_progress.up.sql
new file mode 100644
index 0000000..40d373e
--- /dev/null
+++ b/internal/adapters/store/migrations/020_uploaded_transcripts_progress.up.sql
@@ -0,0 +1,15 @@
+-- Migration 020: where an upload job is, and whether the service is still reporting (ADR-035).
+--
+-- progress    the job's ordered phases as JSON (key, label, state, times, done/total).
+-- progress_at when a phase last REALLY changed or advanced.
+-- alive_at    a heartbeat stamped every few seconds by a goroutine independent of any
+--             engine call. The two differ on purpose: alive says the service is still
+--             reporting, progress says work is moving, and a page that cannot tell a
+--             crashed service from a slow model is the failure this exists to end.
+--
+-- All three are NULL until the job starts. They sit on the already-FORCE'd, RLS-scoped
+-- uploaded_transcripts, so the exact-set isolation test is unaffected.
+ALTER TABLE uploaded_transcripts
+    ADD COLUMN progress    JSONB,
+    ADD COLUMN progress_at TIMESTAMPTZ,
+    ADD COLUMN alive_at    TIMESTAMPTZ;