diff --git a/internal/db/dbq/discover.sql.go b/internal/db/dbq/discover.sql.go
new file mode 100644
index 00000000..3b9ad897
--- /dev/null
+++ b/internal/db/dbq/discover.sql.go
@@ -0,0 +1,207 @@
+// Code generated by sqlc. DO NOT EDIT.
+// versions:
+//   sqlc v1.31.1
+// source: discover.sql
+
+package dbq
+
+import (
+	"context"
+
+	"github.com/jackc/pgx/v5/pgtype"
+)
+
+const listCrossUserLikedTracksForDiscover = `-- name: ListCrossUserLikedTracksForDiscover :many
+SELECT DISTINCT t.id, t.album_id, t.artist_id
+  FROM general_likes gl
+  JOIN tracks t ON t.id = gl.track_id
+ WHERE gl.user_id != $1
+   AND NOT EXISTS (
+       SELECT 1 FROM play_events pe
+        WHERE pe.user_id = $1
+          AND pe.track_id = t.id
+          AND pe.was_skipped = false
+   )
+   AND NOT EXISTS (
+       SELECT 1 FROM general_likes ml
+        WHERE ml.user_id = $1 AND ml.track_id = t.id
+   )
+   AND NOT EXISTS (
+       SELECT 1 FROM lidarr_quarantine q
+        WHERE q.user_id = $1 AND q.track_id = t.id
+   )
+ ORDER BY md5(t.id::text || $2::text)
+ LIMIT 60
+`
+
+type ListCrossUserLikedTracksForDiscoverParams struct {
+	UserID  pgtype.UUID
+	Column2 string
+}
+
+type ListCrossUserLikedTracksForDiscoverRow struct {
+	ID       pgtype.UUID
+	AlbumID  pgtype.UUID
+	ArtistID pgtype.UUID
+}
+
+// Tracks any OTHER user has liked, that this user hasn't played,
+// liked, or had quarantined. On single-user servers this returns 0
+// rows (the caller's slot redistribution rolls the deficit into the
+// other two buckets).
+// $1 = user_id, $2 = date string for md5 ordering.
+func (q *Queries) ListCrossUserLikedTracksForDiscover(ctx context.Context, arg ListCrossUserLikedTracksForDiscoverParams) ([]ListCrossUserLikedTracksForDiscoverRow, error) {
+	rows, err := q.db.Query(ctx, listCrossUserLikedTracksForDiscover, arg.UserID, arg.Column2)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	var items []ListCrossUserLikedTracksForDiscoverRow
+	for rows.Next() {
+		var i ListCrossUserLikedTracksForDiscoverRow
+		if err := rows.Scan(&i.ID, &i.AlbumID, &i.ArtistID); err != nil {
+			return nil, err
+		}
+		items = append(items, i)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+	return items, nil
+}
+
+const listDormantArtistTracksForDiscover = `-- name: ListDormantArtistTracksForDiscover :many
+
+WITH user_artist_counts AS (
+    SELECT t.artist_id, COUNT(*) AS play_count
+      FROM play_events pe
+      JOIN tracks t ON t.id = pe.track_id
+     WHERE pe.user_id = $1
+       AND pe.was_skipped = false
+     GROUP BY t.artist_id
+),
+dormant_artists AS (
+    SELECT a.id
+      FROM artists a
+      LEFT JOIN user_artist_counts uac ON uac.artist_id = a.id
+     WHERE COALESCE(uac.play_count, 0) < 10
+)
+SELECT t.id, t.album_id, t.artist_id
+  FROM tracks t
+  JOIN dormant_artists da ON da.id = t.artist_id
+ WHERE NOT EXISTS (
+       SELECT 1 FROM play_events pe
+        WHERE pe.user_id = $1
+          AND pe.track_id = t.id
+          AND pe.was_skipped = false
+   )
+   AND NOT EXISTS (
+       SELECT 1 FROM general_likes gl
+        WHERE gl.user_id = $1 AND gl.track_id = t.id
+   )
+   AND NOT EXISTS (
+       SELECT 1 FROM lidarr_quarantine q
+        WHERE q.user_id = $1 AND q.track_id = t.id
+   )
+ ORDER BY md5(t.id::text || $2::text)
+ LIMIT 80
+`
+
+type ListDormantArtistTracksForDiscoverParams struct {
+	UserID  pgtype.UUID
+	Column2 string
+}
+
+type ListDormantArtistTracksForDiscoverRow struct {
+	ID       pgtype.UUID
+	AlbumID  pgtype.UUID
+	ArtistID pgtype.UUID
+}
+
+// Discover playlist bucket queries. Each returns up to its LIMIT
+// count of (track_id, album_id, artist_id) triples ordered
+// daily-deterministically via md5(track_id || dateStr). The Go-side
+// bucket allocator (internal/playlists/discover.go) applies
+// per-album (<=2) and per-artist (<=3) caps then redistributes any
+// bucket deficit equally across the others. LIMIT values are
+// generous so the caps + redistribution have headroom.
+// Tracks whose artist this user has played fewer than 10 times total.
+// The artist threshold defines "dormant" — we want to surface the
+// long tail of artists in the user's library that they rarely listen
+// to. Excludes tracks the user has played, liked, or has quarantined.
+// $1 = user_id, $2 = date string for md5 ordering.
+func (q *Queries) ListDormantArtistTracksForDiscover(ctx context.Context, arg ListDormantArtistTracksForDiscoverParams) ([]ListDormantArtistTracksForDiscoverRow, error) {
+	rows, err := q.db.Query(ctx, listDormantArtistTracksForDiscover, arg.UserID, arg.Column2)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	var items []ListDormantArtistTracksForDiscoverRow
+	for rows.Next() {
+		var i ListDormantArtistTracksForDiscoverRow
+		if err := rows.Scan(&i.ID, &i.AlbumID, &i.ArtistID); err != nil {
+			return nil, err
+		}
+		items = append(items, i)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+	return items, nil
+}
+
+const listRandomUnheardTracksForDiscover = `-- name: ListRandomUnheardTracksForDiscover :many
+SELECT t.id, t.album_id, t.artist_id
+  FROM tracks t
+ WHERE NOT EXISTS (
+       SELECT 1 FROM play_events pe
+        WHERE pe.user_id = $1
+          AND pe.track_id = t.id
+          AND pe.was_skipped = false
+   )
+   AND NOT EXISTS (
+       SELECT 1 FROM general_likes gl
+        WHERE gl.user_id = $1 AND gl.track_id = t.id
+   )
+   AND NOT EXISTS (
+       SELECT 1 FROM lidarr_quarantine q
+        WHERE q.user_id = $1 AND q.track_id = t.id
+   )
+ ORDER BY md5(t.id::text || $2::text)
+ LIMIT 200
+`
+
+type ListRandomUnheardTracksForDiscoverParams struct {
+	UserID  pgtype.UUID
+	Column2 string
+}
+
+type ListRandomUnheardTracksForDiscoverRow struct {
+	ID       pgtype.UUID
+	AlbumID  pgtype.UUID
+	ArtistID pgtype.UUID
+}
+
+// Random sample from the entire library. Same exclusion filters as
+// the other two buckets. The generous LIMIT (200) gives the caller
+// headroom for the per-album/artist caps and the slot redistribution.
+// $1 = user_id, $2 = date string for md5 ordering.
+func (q *Queries) ListRandomUnheardTracksForDiscover(ctx context.Context, arg ListRandomUnheardTracksForDiscoverParams) ([]ListRandomUnheardTracksForDiscoverRow, error) {
+	rows, err := q.db.Query(ctx, listRandomUnheardTracksForDiscover, arg.UserID, arg.Column2)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	var items []ListRandomUnheardTracksForDiscoverRow
+	for rows.Next() {
+		var i ListRandomUnheardTracksForDiscoverRow
+		if err := rows.Scan(&i.ID, &i.AlbumID, &i.ArtistID); err != nil {
+			return nil, err
+		}
+		items = append(items, i)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+	return items, nil
+}
diff --git a/internal/db/migrations/0021_discover_variant.down.sql b/internal/db/migrations/0021_discover_variant.down.sql
new file mode 100644
index 00000000..d93aa89b
--- /dev/null
+++ b/internal/db/migrations/0021_discover_variant.down.sql
@@ -0,0 +1,21 @@
+-- Restore the prior tighter CHECK constraints (pre-discover).
+-- This will fail if any rows currently have system_variant = 'discover';
+-- operator must clear those rows manually before rolling back.
+
+ALTER TABLE playlists DROP CONSTRAINT IF EXISTS playlists_kind_variant_consistent;
+ALTER TABLE playlists
+    ADD CONSTRAINT playlists_kind_variant_consistent CHECK (
+        (kind = 'user'   AND system_variant IS NULL)
+        OR
+        (kind = 'system' AND system_variant IN ('for_you', 'songs_like_artist'))
+    );
+
+ALTER TABLE playlists DROP CONSTRAINT IF EXISTS playlists_seed_consistent;
+ALTER TABLE playlists
+    ADD CONSTRAINT playlists_seed_consistent CHECK (
+        (system_variant IS NULL              AND seed_artist_id IS NULL)
+        OR
+        (system_variant = 'for_you'          AND seed_artist_id IS NULL)
+        OR
+        (system_variant = 'songs_like_artist' AND seed_artist_id IS NOT NULL)
+    );
diff --git a/internal/db/migrations/0021_discover_variant.up.sql b/internal/db/migrations/0021_discover_variant.up.sql
new file mode 100644
index 00000000..79af4416
--- /dev/null
+++ b/internal/db/migrations/0021_discover_variant.up.sql
@@ -0,0 +1,27 @@
+-- Adds 'discover' as an accepted system_variant alongside 'for_you'
+-- and 'songs_like_artist'. Discover has no seed_artist_id (like
+-- for_you), so the variant->seed coherence check is extended to
+-- accept (system_variant = 'discover' AND seed_artist_id IS NULL).
+--
+-- Drops + re-adds both CHECK constraints; PostgreSQL doesn't have
+-- ALTER CONSTRAINT for CHECK so this is the standard pattern.
+
+ALTER TABLE playlists DROP CONSTRAINT IF EXISTS playlists_kind_variant_consistent;
+ALTER TABLE playlists
+    ADD CONSTRAINT playlists_kind_variant_consistent CHECK (
+        (kind = 'user'   AND system_variant IS NULL)
+        OR
+        (kind = 'system' AND system_variant IN ('for_you', 'songs_like_artist', 'discover'))
+    );
+
+ALTER TABLE playlists DROP CONSTRAINT IF EXISTS playlists_seed_consistent;
+ALTER TABLE playlists
+    ADD CONSTRAINT playlists_seed_consistent CHECK (
+        (system_variant IS NULL                 AND seed_artist_id IS NULL)
+        OR
+        (system_variant = 'for_you'             AND seed_artist_id IS NULL)
+        OR
+        (system_variant = 'discover'            AND seed_artist_id IS NULL)
+        OR
+        (system_variant = 'songs_like_artist'   AND seed_artist_id IS NOT NULL)
+    );
diff --git a/internal/db/queries/discover.sql b/internal/db/queries/discover.sql
new file mode 100644
index 00000000..8d13392f
--- /dev/null
+++ b/internal/db/queries/discover.sql
@@ -0,0 +1,98 @@
+-- Discover playlist bucket queries. Each returns up to its LIMIT
+-- count of (track_id, album_id, artist_id) triples ordered
+-- daily-deterministically via md5(track_id || dateStr). The Go-side
+-- bucket allocator (internal/playlists/discover.go) applies
+-- per-album (<=2) and per-artist (<=3) caps then redistributes any
+-- bucket deficit equally across the others. LIMIT values are
+-- generous so the caps + redistribution have headroom.
+
+-- name: ListDormantArtistTracksForDiscover :many
+-- Tracks whose artist this user has played fewer than 10 times total.
+-- The artist threshold defines "dormant" — we want to surface the
+-- long tail of artists in the user's library that they rarely listen
+-- to. Excludes tracks the user has played, liked, or has quarantined.
+-- $1 = user_id, $2 = date string for md5 ordering.
+WITH user_artist_counts AS (
+    SELECT t.artist_id, COUNT(*) AS play_count
+      FROM play_events pe
+      JOIN tracks t ON t.id = pe.track_id
+     WHERE pe.user_id = $1
+       AND pe.was_skipped = false
+     GROUP BY t.artist_id
+),
+dormant_artists AS (
+    SELECT a.id
+      FROM artists a
+      LEFT JOIN user_artist_counts uac ON uac.artist_id = a.id
+     WHERE COALESCE(uac.play_count, 0) < 10
+)
+SELECT t.id, t.album_id, t.artist_id
+  FROM tracks t
+  JOIN dormant_artists da ON da.id = t.artist_id
+ WHERE NOT EXISTS (
+       SELECT 1 FROM play_events pe
+        WHERE pe.user_id = $1
+          AND pe.track_id = t.id
+          AND pe.was_skipped = false
+   )
+   AND NOT EXISTS (
+       SELECT 1 FROM general_likes gl
+        WHERE gl.user_id = $1 AND gl.track_id = t.id
+   )
+   AND NOT EXISTS (
+       SELECT 1 FROM lidarr_quarantine q
+        WHERE q.user_id = $1 AND q.track_id = t.id
+   )
+ ORDER BY md5(t.id::text || $2::text)
+ LIMIT 80;
+
+-- name: ListCrossUserLikedTracksForDiscover :many
+-- Tracks any OTHER user has liked, that this user hasn't played,
+-- liked, or had quarantined. On single-user servers this returns 0
+-- rows (the caller's slot redistribution rolls the deficit into the
+-- other two buckets).
+-- $1 = user_id, $2 = date string for md5 ordering.
+SELECT DISTINCT t.id, t.album_id, t.artist_id
+  FROM general_likes gl
+  JOIN tracks t ON t.id = gl.track_id
+ WHERE gl.user_id != $1
+   AND NOT EXISTS (
+       SELECT 1 FROM play_events pe
+        WHERE pe.user_id = $1
+          AND pe.track_id = t.id
+          AND pe.was_skipped = false
+   )
+   AND NOT EXISTS (
+       SELECT 1 FROM general_likes ml
+        WHERE ml.user_id = $1 AND ml.track_id = t.id
+   )
+   AND NOT EXISTS (
+       SELECT 1 FROM lidarr_quarantine q
+        WHERE q.user_id = $1 AND q.track_id = t.id
+   )
+ ORDER BY md5(t.id::text || $2::text)
+ LIMIT 60;
+
+-- name: ListRandomUnheardTracksForDiscover :many
+-- Random sample from the entire library. Same exclusion filters as
+-- the other two buckets. The generous LIMIT (200) gives the caller
+-- headroom for the per-album/artist caps and the slot redistribution.
+-- $1 = user_id, $2 = date string for md5 ordering.
+SELECT t.id, t.album_id, t.artist_id
+  FROM tracks t
+ WHERE NOT EXISTS (
+       SELECT 1 FROM play_events pe
+        WHERE pe.user_id = $1
+          AND pe.track_id = t.id
+          AND pe.was_skipped = false
+   )
+   AND NOT EXISTS (
+       SELECT 1 FROM general_likes gl
+        WHERE gl.user_id = $1 AND gl.track_id = t.id
+   )
+   AND NOT EXISTS (
+       SELECT 1 FROM lidarr_quarantine q
+        WHERE q.user_id = $1 AND q.track_id = t.id
+   )
+ ORDER BY md5(t.id::text || $2::text)
+ LIMIT 200;