diff --git a/supabase/migrations/20260717120000_corrector_public_titles_only.sql b/supabase/migrations/20260717120000_corrector_public_titles_only.sql
new file mode 100644
index 000000000..46d062920
--- /dev/null
+++ b/supabase/migrations/20260717120000_corrector_public_titles_only.sql
@@ -0,0 +1,87 @@
+-- Scope the clinical query-term corrector's vocabulary to the public corpus.
+--
+-- correct_clinical_query_terms() is SECURITY DEFINER and therefore bypasses RLS.
+-- It previously built its spell-correction vocabulary from EVERY indexed document
+-- title AND every enabled rag_aliases row regardless of owner, while the rest of
+-- retrieval is strictly owner-scoped and fail-closed (see
+-- docs/tenancy-defense-in-depth-review.md and retrieval_owner_matches). Both
+-- rag_aliases (owner_id column; deep-memory.ts persists owner-scoped aliases and
+-- canonicals for private documents) and documents carry private rows, so the unscoped
+-- reads let a private tenant's terms bias — and, via observable query rewriting, leak
+-- the existence of — another tenant's private aliases/titles: a cross-tenant side-channel.
+--
+-- Fix: restrict every vocabulary source (both alias selects and the title scan) to the
+-- shared public corpus (owner_id is null). Signature is unchanged, so callers and
+-- generated types need no update. Idempotent CREATE OR REPLACE.
+--
+-- NOTE ON SEQUENCING: this migration is the forward change. supabase/schema.sql (the
+-- canonical replay reference) and supabase/drift-manifest.json must be updated to
+-- match when this migration is applied to the live project — both require the Docker
+-- replay (`npm run drift:manifest`) that is unavailable in the CI/agent sandbox. Apply
+-- this migration + the schema.sql mirror + the manifest regen together as one operator
+-- step (live apply is confirmation-gated). Until applied, live == schema.sql == manifest
+-- (all the pre-scope version), so no drift is introduced by merging this file alone.
+
+CREATE OR REPLACE FUNCTION public.correct_clinical_query_terms(input_query text, min_sim real DEFAULT 0.45)
+ RETURNS text
+ LANGUAGE plpgsql
+ STABLE SECURITY DEFINER
+ SET search_path TO 'public', 'extensions', 'pg_temp'
+AS $function$
+declare
+  vocab text[];
+  tokens text[];
+  tok text;
+  best text;
+  best_sim real;
+  corrected text[] := array[]::text[];
+  changed boolean := false;
+begin
+  if input_query is null or length(trim(input_query)) = 0 then
+    return input_query;
+  end if;
+
+  -- Build the known-term vocabulary once per call. Every source is scoped to the
+  -- public (null-owner) corpus: rag_aliases carries an owner_id (deep-memory persists
+  -- owner-scoped aliases/canonicals for private documents), so an unscoped alias read
+  -- would leak private-document-derived terms across tenants just like the title scan.
+  select array_agg(distinct term) into vocab
+  from (
+    select lower(alias) as term from public.rag_aliases where enabled and owner_id is null and length(alias) between 4 and 40
+    union
+    select lower(canonical) from public.rag_aliases where enabled and owner_id is null and length(canonical) between 4 and 40
+    union
+    select w from public.documents d, lateral unnest(regexp_split_to_array(lower(d.title), '[^a-z]+')) as w
+    -- Public (null-owner) titles only: keep the correction vocabulary tenant-safe.
+    where d.status = 'indexed' and d.owner_id is null and length(w) between 4 and 40
+  ) t;
+
+  tokens := regexp_split_to_array(lower(input_query), '\s+');
+  foreach tok in array tokens loop
+    if length(tok) < 4 or tok = any(vocab) then
+      corrected := corrected || tok;
+      continue;
+    end if;
+    best := null;
+    best_sim := 0;
+    select v, similarity(v, tok) into best, best_sim
+    from unnest(vocab) as v
+    order by similarity(v, tok) desc
+    limit 1;
+    if best is not null and best_sim >= min_sim and best <> tok and length(best) >= length(tok) then
+      corrected := corrected || best;
+      changed := true;
+    else
+      corrected := corrected || tok;
+    end if;
+  end loop;
+
+  if not changed then
+    return input_query;
+  end if;
+  return array_to_string(corrected, ' ');
+end;
+$function$;
+
+revoke execute on function public.correct_clinical_query_terms(text, real) from public, anon, authenticated;
+grant execute on function public.correct_clinical_query_terms(text, real) to service_role;
diff --git a/tests/supabase-schema.test.ts b/tests/supabase-schema.test.ts
index 6de21304a..2e1b4634a 100644
--- a/tests/supabase-schema.test.ts
+++ b/tests/supabase-schema.test.ts
@@ -1187,3 +1187,55 @@ describe("Supabase Preview replay guards", () => {
     expect(ingestionRpcPrivilegesDuplicateMigration).toContain("select 1 where false;");
   });
 });
+
+describe("Clinical query-term corrector — tenant-safe vocabulary (F10)", () => {
+  // The fix ships as a forward migration; schema.sql + drift-manifest are synced at
+  // the Docker-gated apply step, so this asserts the migration rather than schema.sql.
+  const correctorPublicTitlesMigration = readFileSync(
+    new URL("../supabase/migrations/20260717120000_corrector_public_titles_only.sql", import.meta.url),
+    "utf8",
+  )
+    .replace(/\s+/g, " ")
+    .toLowerCase();
+
+  it("recreates the corrector scoped to the public (null-owner) title corpus", () => {
+    expect(correctorPublicTitlesMigration).toContain(
+      "create or replace function public.correct_clinical_query_terms(input_query text, min_sim real default 0.45)",
+    );
+    // SECURITY DEFINER bypasses RLS, so the title scan must be owner-scoped to keep a
+    // private tenant's title tokens out of every caller's correction vocabulary.
+    expect(correctorPublicTitlesMigration).toContain(
+      "where d.status = 'indexed' and d.owner_id is null and length(w) between 4 and 40",
+    );
+    // Regression guard: the old unscoped predicate must not survive in the migration.
+    expect(correctorPublicTitlesMigration).not.toContain("where d.status = 'indexed' and length(w) between 4 and 40");
+  });
+
+  it("scopes the rag_aliases vocabulary sources to public (null-owner) rows", () => {
+    // rag_aliases carries an owner_id (deep-memory persists owner-scoped aliases for
+    // private documents), so both alias reads must be owner-scoped too — otherwise the
+    // title fix alone still leaks private-document-derived terms across tenants.
+    expect(correctorPublicTitlesMigration).toContain(
+      "select lower(alias) as term from public.rag_aliases where enabled and owner_id is null and length(alias) between 4 and 40",
+    );
+    expect(correctorPublicTitlesMigration).toContain(
+      "select lower(canonical) from public.rag_aliases where enabled and owner_id is null and length(canonical) between 4 and 40",
+    );
+    // Regression guard: the old unscoped alias reads must not survive.
+    expect(correctorPublicTitlesMigration).not.toContain(
+      "from public.rag_aliases where enabled and length(alias) between 4 and 40",
+    );
+    expect(correctorPublicTitlesMigration).not.toContain(
+      "from public.rag_aliases where enabled and length(canonical) between 4 and 40",
+    );
+  });
+
+  it("keeps the corrector execute privilege confined to service_role", () => {
+    expect(correctorPublicTitlesMigration).toContain(
+      "revoke execute on function public.correct_clinical_query_terms(text, real) from public, anon, authenticated;",
+    );
+    expect(correctorPublicTitlesMigration).toContain(
+      "grant execute on function public.correct_clinical_query_terms(text, real) to service_role;",
+    );
+  });
+});