feat(derived): encrypt flashcards, study guides, room summaries (#518)

CLAUDE.md

@@ -106,7 +106,7 @@ make explore # Chapter 2: bounded AI exploration of the ru
 
 ## Gotchas
 
-- Column-level encryption is on for sensitive columns: `user_profiles.name`/`first_name`/`last_name`/`bio`/`location` (these moved off `users` to `user_profiles` in the 0024 identity split), Google OAuth tokens, `messages.content`, `room_messages.text`, `sessions.summary_json`, `documents.summary` + `concept_notes` + `extracted_text` (the RAG OCR text added in 0030), `notes.title`/`body`/`last_summary`, `assignments.notes`/`points_possible`/`points_earned` (the enrollment-keyed gradebook table; points columns carry numeric semantics — use `decrypt_numeric` at read), `feedback.comment`/`topic` + `issue_reports.topic`/`description` (free-text user input, #520), and `quiz_attempts.questions_json`/`answers_json` + `quiz_context.context_json` (quiz performance data, #521; scalar analytics columns — score/total/difficulty/completed_at — stay plaintext). Helpers live in `backend/services/encryption.py`; use `encrypt_if_present` at write boundaries and `decrypt_if_present` / `decrypt_numeric` at read boundaries (including before injecting into AI prompts). `ENCRYPTION_KEY` must be set (32 bytes as 64 hex chars; generate via `python -c "import secrets; print(secrets.token_hex(32))"`). Deliberate exception: `newsletter_emails.email` stays plaintext (ADR 0026) — the UNIQUE constraint, lookup index, and both subscribe/allowlist upserts key on the value, and AES-GCM's per-call nonce breaks value equality.
+- Column-level encryption is on for sensitive columns: `user_profiles.name`/`first_name`/`last_name`/`bio`/`location` (these moved off `users` to `user_profiles` in the 0024 identity split), Google OAuth tokens, `messages.content`, `room_messages.text`, `sessions.summary_json`, `documents.summary` + `concept_notes` + `extracted_text` (the RAG OCR text added in 0030), `notes.title`/`body`/`last_summary`, `assignments.notes`/`points_possible`/`points_earned` (the enrollment-keyed gradebook table; points columns carry numeric semantics — use `decrypt_numeric` at read), `feedback.comment`/`topic` + `issue_reports.topic`/`description` (free-text user input, #520), `quiz_attempts.questions_json`/`answers_json` + `quiz_context.context_json` (quiz performance data, #521; scalar analytics columns — score/total/difficulty/completed_at — stay plaintext), and `flashcards.front`/`back` + `study_guides.content` + `room_summaries.summary` (derived content, #518; `study_guides.content` uses the JSON pair, `room_summaries.summary` keys its cache on the separate plaintext `member_hash` column so encryption doesn't affect cache-hit lookups). Helpers live in `backend/services/encryption.py`; use `encrypt_if_present` at write boundaries and `decrypt_if_present` / `decrypt_numeric` at read boundaries (including before injecting into AI prompts). `ENCRYPTION_KEY` must be set (32 bytes as 64 hex chars; generate via `python -c "import secrets; print(secrets.token_hex(32))"`). Deliberate exception: `newsletter_emails.email` stays plaintext (ADR 0026) — the UNIQUE constraint, lookup index, and both subscribe/allowlist upserts key on the value, and AES-GCM's per-call nonce breaks value equality.
 - Knowledge-graph mastery is now an append-only `node_mastery_events` table (replaced the `graph_nodes.mastery_events` JSON column in 0023); node/edge dedup is enforced by UNIQUE constraints. Don't read/write a `mastery_events` column.
 - Optional cross-worker cache (#97): `services/cache.py` wraps Redis and is **off by default** — with no `REDIS_URL` set it's a zero-overhead no-op and never fails a request (any Redis error → clean miss + warning). Currently backs the content-addressed OCR/extraction cache (`extraction_service.extract_text_from_file`, keyed on `sha256(file_bytes)` + engine). The `redis` dependency is only imported when `REDIS_URL` is set.
 - HTTP caching (#99): conditional GETs use `services/http_cache.py` (`make_etag`/`conditional`/`cached_json`). `Cache-Control` on these routes is **always `private`, never `public`** — the responses carry user-scoped, app-decrypted columns that must never be cached at a shared proxy/CDN. Derive the ETag from cheap change-keys (ids, `updated_at`, existing content hashes), not from the fully-built payload.

backend/db/backfill_encryption.py

@@ -302,6 +302,18 @@ def backfill_quiz_context(apply: bool) -> dict:
  return _encrypt_json_column("quiz_context", "context_json", pk="id", apply=apply)
 
 
+def backfill_flashcards(apply: bool) -> dict:
+ return _encrypt_text_column("flashcards", ["front", "back"], pk="id", apply=apply)
+
+
+def backfill_study_guides(apply: bool) -> dict:
+ return _encrypt_json_column("study_guides", "content", pk="id", apply=apply)
+
+
+def backfill_room_summaries(apply: bool) -> dict:
+ return _encrypt_text_column("room_summaries", ["summary"], pk="room_id", apply=apply)
+
+
 RUNNERS: dict[str, Callable[[bool], dict]] = {
  "users": backfill_users,
  "user_settings": backfill_user_settings,
@@ -315,6 +327,9 @@ def backfill_quiz_context(apply: bool) -> dict:
  "issue_reports": backfill_issue_reports,
  "quiz_attempts": backfill_quiz_attempts,
  "quiz_context": backfill_quiz_context,
+ "flashcards": backfill_flashcards,
+ "study_guides": backfill_study_guides,
+ "room_summaries": backfill_room_summaries,
 }
 
 

backend/db/seed_local_rich.py

@@ -553,7 +553,7 @@ def seed_notes_documents() -> None:
  )
 
 
-# (fc_id, user_id, offering_id, topic, front, back) — plaintext, grouped by topic.
+# (fc_id, user_id, offering_id, topic, front, back) — 🔒 front/back, grouped by topic.
 _FLASHCARDS = [
  ("rich-fc-cs-1", USER_ACTIVE, OFF_CS_F25, "CS Basics",
  "What is a variable?", "A named storage location for a value."),
@@ -579,8 +579,9 @@ def seed_flashcards() -> None:
  "user_id": user_id,
  "offering_id": off_id,
  "topic": topic,
- "front": front,
- "back": back,
+ # 🔒 front / back (#518)
+ "front": encrypt_if_present(front),
+ "back": encrypt_if_present(back),
  },
  )
 
@@ -625,11 +626,26 @@ def seed_study_guides() -> None:
  "offering_id": off_id,
  "exam_id": exam_id,
  "generated_at": generated_at,
- "content": content,
+ # 🔒 content (#518)
+ "content": encrypt_json(content),
  },
  )
 
 
+def seed_room_summaries() -> None:
+ # #518: room_summaries.summary is 🔒. PK is room_id (no id column), so this
+ # can't go through insert_if_absent.
+ if not table("room_summaries").select("room_id", filters={"room_id": f"eq.{ROOM_STUDY}"}):
+ table("room_summaries").insert({
+ "room_id": ROOM_STUDY,
+ "summary": encrypt_if_present("The group is reviewing recursion before the midterm."),
+ "member_hash": "rich-member-hash-v1",
+ })
+ h.record("room_summaries", created=True)
+ else:
+ h.record("room_summaries", created=False)
+
+
 # (qa_id, concept_node_id, difficulty, score, total, questions_json, answers_json, completed_at)
 _QUIZ_ATTEMPTS = [
  ("rich-qa-cs-variables-1", "rich-node-cs-variables", "easy", 9, 10,
@@ -766,6 +782,7 @@ def seed_sessions() -> None:
  "schools", "courses", "course_offerings", "users", "user_profiles", "user_roles",
  "enrollments", "graph_nodes", "graph_edges", "node_mastery_events",
  "gradebook_categories", "assignments", "rooms", "room_members", "room_messages",
+ "room_summaries",
  "notes", "documents", "flashcards", "study_guides", "quiz_attempts", "quiz_context",
  "sessions", "messages", "feedback", "issue_reports",
 ]
@@ -785,6 +802,7 @@ def main() -> None:
  seed_notes_documents()
  seed_flashcards()
  seed_study_guides()
+ seed_room_summaries()
  seed_quiz()
  seed_feedback()
  seed_sessions()

backend/e2e_oracles/gather.py

@@ -195,6 +195,10 @@ def run_counts(args: argparse.Namespace) -> tuple[list[Finding], int]:
  ("quiz_attempts", "id", "questions_json"),
  ("quiz_attempts", "id", "answers_json"),
  ("quiz_context", "id", "context_json"),
+ ("flashcards", "id", "front"),
+ ("flashcards", "id", "back"),
+ ("study_guides", "id", "content"),
+ ("room_summaries", "room_id", "summary"),
 )
 
 

backend/routes/flashcards.py

@@ -14,7 +14,7 @@
 from services.academics import resolve_offering, term_id_for_label
 from services.auth_guard import require_self, get_session_user_id
 from services.achievement_service import check_achievements
-from services.encryption import decrypt_if_present, decrypt_json
+from services.encryption import decrypt_if_present, decrypt_json, encrypt_if_present
 from services.flashcard_import_service import (
  dedup_against_existing,
  check_rate_limit,
@@ -236,8 +236,8 @@ def generate(body: GenerateFlashcardsBody, request: Request):
  "id": str(uuid.uuid4()),
  "user_id": body.user_id,
  "topic": body.topic,
- "front": c["front"],
- "back": c["back"],
+ "front": encrypt_if_present(c["front"]),
+ "back": encrypt_if_present(c["back"]),
  "times_reviewed": 0,
  "last_reviewed_at": None,
  "created_at": now,
@@ -261,8 +261,16 @@ def generate(body: GenerateFlashcardsBody, request: Request):
  except Exception:
  pass
 
+ # rows_to_insert holds the ciphertext just written; the response the
+ # frontend renders (freshly generated cards, before any list re-fetch)
+ # must carry plaintext front/back, not the encrypted insert payload.
+ response_cards = [
+ {**row, "front": decrypt_if_present(row["front"]), "back": decrypt_if_present(row["back"])}
+ for row in rows_to_insert
+ ]
+
  return {
- "flashcards": rows_to_insert,
+ "flashcards": response_cards,
  "context_used": {
  "documents_found": len(documents),
  "weak_concepts_found": len(weak_concepts),
@@ -291,6 +299,9 @@ def get_flashcards(
  "id,user_id,topic,offering_id,front,back,times_reviewed,last_rating,last_reviewed_at,created_at",
  filters=filters, order="created_at.desc"
  ) or []
+ for r in rows:
+ r["front"] = decrypt_if_present(r.get("front"))
+ r["back"] = decrypt_if_present(r.get("back"))
  if semester:
  # Term scoping (#141). This route is user-wide (no course id), so
  # the filter works on the cards' offering: cards from the selected
@@ -397,8 +408,8 @@ def import_commit(body: ImportCommitBody, request: Request):
  "user_id": body.user_id,
  "topic": body.topic,
  "offering_id": offering_id,
- "front": c["front"],
- "back": c["back"],
+ "front": encrypt_if_present(c["front"]),
+ "back": encrypt_if_present(c["back"]),
  "times_reviewed": 0,
  "last_reviewed_at": None,
  "created_at": now,

backend/routes/study_guide.py

@@ -4,6 +4,7 @@
 Study guide generation and caching.
 """
 
+import logging
 import uuid
 from datetime import datetime, timezone
 
@@ -24,10 +25,17 @@
 )
 from services.graph_service import get_courses as graph_get_courses
 from services.auth_guard import require_self
-from services.encryption import decrypt_if_present, decrypt_json
+from services.encryption import (
+ decrypt_if_present,
+ decrypt_json,
+ decrypt_json_column,
+ encrypt_json,
+)
 from services.http_cache import cached_json, conditional, make_etag
 from services.request_context import current_request_id
 
+logger = logging.getLogger(__name__)
+
 router = APIRouter()
 
 
@@ -198,7 +206,7 @@ def _generate_and_insert(user_id: str, offering_id: str, exam_id: str) -> dict:
  "offering_id": offering_id,
  "exam_id": exam_id,
  "generated_at": now,
- "content": content,
+ "content": encrypt_json(content),
  }
  table("study_guides").insert(row)
 
@@ -245,7 +253,14 @@ def get_cached_guides(user_id: str, request: Request):
 
  result = []
  for g in guides:
- content = g.get("content") or {}
+ try:
+ content = decrypt_json_column(g.get("content")) or {}
+ except Exception:
+ logger.warning(
+ "get_cached_guides: content decrypt failed for guide %s; degrading",
+ g.get("id"),
+ )
+ content = {}
  course_id = offering_to_course.get(g.get("offering_id"))
  result.append({
  "id": g["id"],
@@ -330,7 +345,7 @@ def get_guide(
  )
  if cached:
  row = cached[0]
- return {"guide": row["content"], "generated_at": row["generated_at"], "cached": True}
+ return {"guide": decrypt_json_column(row["content"]), "generated_at": row["generated_at"], "cached": True}
 
  result = _generate_and_insert(user_id, offering_id, exam_id)
  return {"guide": result["content"], "generated_at": result["generated_at"], "cached": False}

backend/services/flashcard_import_service.py

@@ -33,6 +33,7 @@
 from agents.flashcard import flashcard_agent
 from agents.usage import record_agent_usage
 from services import extraction_service
+from services.encryption import decrypt_if_present
 
 logger = logging.getLogger(__name__)
 
@@ -76,7 +77,9 @@ def dedup_against_existing(
  filters["topic"] = f"eq.{topic}"
 
  existing = table("flashcards").select("front", filters=filters) or []
- existing_norm = [_normalize(r.get("front", "")) for r in existing]
+ existing_norm = [
+ _normalize(decrypt_if_present(r.get("front", "")) or "") for r in existing
+ ]
 
  keep: list[Card] = []
  skipped: list[Card] = []

backend/services/social_cache_service.py

@@ -12,6 +12,7 @@
 from datetime import datetime, timezone
 
 from db.connection import table
+from services.encryption import decrypt_if_present, encrypt_if_present
 
 
 def _compute_hash(member_summaries: list[str]) -> str:
@@ -26,15 +27,15 @@ def get_cached_summary(room_id: str, member_summaries: list[str]) -> str | None:
  filters={"room_id": f"eq.{room_id}"},
  )
  if rows and rows[0]["member_hash"] == current_hash:
- return rows[0]["summary"]
+ return decrypt_if_present(rows[0]["summary"])
  return None
 
 
 def save_summary(room_id: str, member_summaries: list[str], summary: str) -> None:
  table("room_summaries").upsert(
  {
  "room_id": room_id,
- "summary": summary,
+ "summary": encrypt_if_present(summary),
  "member_hash": _compute_hash(member_summaries),
  "updated_at": datetime.now(timezone.utc).isoformat(),
  },

backend/tests/integration/test_encryption_roundtrip.py

@@ -46,6 +46,13 @@
  "Upload stuck"),
  ("issue_reports.description", "issue_reports", "id", "rich-issue-1",
  "description", "Syllabus upload spins forever."),
+ ("flashcards.front", "flashcards", "id", "rich-fc-cs-1", "front",
+ "What is a variable?"),
+ ("flashcards.back", "flashcards", "id", "rich-fc-cs-1", "back",
+ "A named storage location for a value."),
+ ("room_summaries.summary", "room_summaries", "room_id",
+ "rich-room-study-group", "summary",
+ "The group is reviewing recursion before the midterm."),
 ]
 
 # (label, id_value, column, expected_number)
@@ -123,6 +130,14 @@ def test_quiz_context_context_json_is_ciphertext_and_decrypts(db_conn):
  assert decoded["asked"] == 2
 
 
+def test_study_guides_content_is_ciphertext_and_decrypts(db_conn):
+ """#518: study_guides.content needs the JSON pair (encrypt_json/decrypt_json)."""
+ raw = _raw(db_conn, "study_guides", "id", "rich-guide-cs-f25-mid", "content")
+ assert isinstance(raw, str), "content stored as PLAINTEXT JSONB — encryption regressed"
+ decoded = decrypt_json(raw)
+ assert decoded["exam"] == "Midterm Exam"
+
+
 def test_documents_concept_notes_uses_json_encryption(db_conn):
  """documents.concept_notes is stored via encrypt_json (a list of concepts)."""
  raw = _raw(db_conn, "documents", "id", "rich-doc-cs-syllabus", "concept_notes")

backend/tests/test_flashcard_import_service.py

@@ -59,6 +59,28 @@ def test_filters_by_topic_when_offering_id_is_none(self):
  assert "offering_id" not in str(call_kwargs)
  assert "course_id" not in str(call_kwargs)
 
+ def test_dedup_matches_against_encrypted_existing_cards(self):
+ # #518: dedupe must be decrypt-aware — existing rows are ciphertext.
+ from services.encryption import encrypt
+ with patch("services.flashcard_import_service.table") as t:
+ t.return_value.select.return_value = [
+ {"front": encrypt("What is a variable?")}
+ ]
+ new = [{"front": "What is a variable?", "back": "A named storage location."}]
+ keep, skipped = svc.dedup_against_existing("u1", "c1", new)
+ assert keep == []
+ assert skipped == new
+
+ def test_dedup_still_matches_legacy_plaintext_rows(self):
+ # Pre-backfill rows are still plaintext — decrypt_if_present falls
+ # back to the raw value, so dedupe keeps working unchanged.
+ with patch("services.flashcard_import_service.table") as t:
+ t.return_value.select.return_value = [{"front": "What is a variable?"}]
+ new = [{"front": "What is a variable?", "back": "A named storage location."}]
+ keep, skipped = svc.dedup_against_existing("u1", "c1", new)
+ assert keep == []
+ assert skipped == new
+
 
 # ── check_rate_limit ─────────────────────────────────────────────────────────
 

CLAUDE.md

@@ -106,7 +106,7 @@ make explore # Chapter 2: bounded AI exploration of the ru
 
 ## Gotchas
 
-- Column-level encryption is on for sensitive columns: `user_profiles.name`/`first_name`/`last_name`/`bio`/`location` (these moved off `users` to `user_profiles` in the 0024 identity split), Google OAuth tokens, `messages.content`, `room_messages.text`, `sessions.summary_json`, `documents.summary` + `concept_notes` + `extracted_text` (the RAG OCR text added in 0030), `notes.title`/`body`/`last_summary`, `assignments.notes`/`points_possible`/`points_earned` (the enrollment-keyed gradebook table; points columns carry numeric semantics — use `decrypt_numeric` at read), `feedback.comment`/`topic` + `issue_reports.topic`/`description` (free-text user input, #520), and `quiz_attempts.questions_json`/`answers_json` + `quiz_context.context_json` (quiz performance data, #521; scalar analytics columns — score/total/difficulty/completed_at — stay plaintext). Helpers live in `backend/services/encryption.py`; use `encrypt_if_present` at write boundaries and `decrypt_if_present` / `decrypt_numeric` at read boundaries (including before injecting into AI prompts). `ENCRYPTION_KEY` must be set (32 bytes as 64 hex chars; generate via `python -c "import secrets; print(secrets.token_hex(32))"`). Deliberate exception: `newsletter_emails.email` stays plaintext (ADR 0026) — the UNIQUE constraint, lookup index, and both subscribe/allowlist upserts key on the value, and AES-GCM's per-call nonce breaks value equality.
+- Column-level encryption is on for sensitive columns: `user_profiles.name`/`first_name`/`last_name`/`bio`/`location` (these moved off `users` to `user_profiles` in the 0024 identity split), Google OAuth tokens, `messages.content`, `room_messages.text`, `sessions.summary_json`, `documents.summary` + `concept_notes` + `extracted_text` (the RAG OCR text added in 0030), `notes.title`/`body`/`last_summary`, `assignments.notes`/`points_possible`/`points_earned` (the enrollment-keyed gradebook table; points columns carry numeric semantics — use `decrypt_numeric` at read), `feedback.comment`/`topic` + `issue_reports.topic`/`description` (free-text user input, #520), `quiz_attempts.questions_json`/`answers_json` + `quiz_context.context_json` (quiz performance data, #521; scalar analytics columns — score/total/difficulty/completed_at — stay plaintext), and `flashcards.front`/`back` + `study_guides.content` + `room_summaries.summary` (derived content, #518; `study_guides.content` uses the JSON pair, `room_summaries.summary` keys its cache on the separate plaintext `member_hash` column so encryption doesn't affect cache-hit lookups). Helpers live in `backend/services/encryption.py`; use `encrypt_if_present` at write boundaries and `decrypt_if_present` / `decrypt_numeric` at read boundaries (including before injecting into AI prompts). `ENCRYPTION_KEY` must be set (32 bytes as 64 hex chars; generate via `python -c "import secrets; print(secrets.token_hex(32))"`). Deliberate exception: `newsletter_emails.email` stays plaintext (ADR 0026) — the UNIQUE constraint, lookup index, and both subscribe/allowlist upserts key on the value, and AES-GCM's per-call nonce breaks value equality.
 - Knowledge-graph mastery is now an append-only `node_mastery_events` table (replaced the `graph_nodes.mastery_events` JSON column in 0023); node/edge dedup is enforced by UNIQUE constraints. Don't read/write a `mastery_events` column.
 - Optional cross-worker cache (#97): `services/cache.py` wraps Redis and is **off by default** — with no `REDIS_URL` set it's a zero-overhead no-op and never fails a request (any Redis error → clean miss + warning). Currently backs the content-addressed OCR/extraction cache (`extraction_service.extract_text_from_file`, keyed on `sha256(file_bytes)` + engine). The `redis` dependency is only imported when `REDIS_URL` is set.
 - HTTP caching (#99): conditional GETs use `services/http_cache.py` (`make_etag`/`conditional`/`cached_json`). `Cache-Control` on these routes is **always `private`, never `public`** — the responses carry user-scoped, app-decrypted columns that must never be cached at a shared proxy/CDN. Derive the ETag from cheap change-keys (ids, `updated_at`, existing content hashes), not from the fully-built payload.

backend/db/backfill_encryption.py

@@ -302,6 +302,18 @@ def backfill_quiz_context(apply: bool) -> dict:
  return _encrypt_json_column("quiz_context", "context_json", pk="id", apply=apply)
 
 
+def backfill_flashcards(apply: bool) -> dict:
+ return _encrypt_text_column("flashcards", ["front", "back"], pk="id", apply=apply)
+
+
+def backfill_study_guides(apply: bool) -> dict:
+ return _encrypt_json_column("study_guides", "content", pk="id", apply=apply)
+
+
+def backfill_room_summaries(apply: bool) -> dict:
+ return _encrypt_text_column("room_summaries", ["summary"], pk="room_id", apply=apply)
+
+
 RUNNERS: dict[str, Callable[[bool], dict]] = {
  "users": backfill_users,
  "user_settings": backfill_user_settings,
@@ -315,6 +327,9 @@ def backfill_quiz_context(apply: bool) -> dict:
  "issue_reports": backfill_issue_reports,
  "quiz_attempts": backfill_quiz_attempts,
  "quiz_context": backfill_quiz_context,
+ "flashcards": backfill_flashcards,
+ "study_guides": backfill_study_guides,
+ "room_summaries": backfill_room_summaries,
 }
 
 

backend/db/seed_local_rich.py

@@ -553,7 +553,7 @@ def seed_notes_documents() -> None:
  )
 
 
-# (fc_id, user_id, offering_id, topic, front, back) — plaintext, grouped by topic.
+# (fc_id, user_id, offering_id, topic, front, back) — 🔒 front/back, grouped by topic.
 _FLASHCARDS = [
  ("rich-fc-cs-1", USER_ACTIVE, OFF_CS_F25, "CS Basics",
  "What is a variable?", "A named storage location for a value."),
@@ -579,8 +579,9 @@ def seed_flashcards() -> None:
  "user_id": user_id,
  "offering_id": off_id,
  "topic": topic,
- "front": front,
- "back": back,
+ # 🔒 front / back (#518)
+ "front": encrypt_if_present(front),
+ "back": encrypt_if_present(back),
  },
  )
 
@@ -625,11 +626,26 @@ def seed_study_guides() -> None:
  "offering_id": off_id,
  "exam_id": exam_id,
  "generated_at": generated_at,
- "content": content,
+ # 🔒 content (#518)
+ "content": encrypt_json(content),
  },
  )
 
 
+def seed_room_summaries() -> None:
+ # #518: room_summaries.summary is 🔒. PK is room_id (no id column), so this
+ # can't go through insert_if_absent.
+ if not table("room_summaries").select("room_id", filters={"room_id": f"eq.{ROOM_STUDY}"}):
+ table("room_summaries").insert({
+ "room_id": ROOM_STUDY,
+ "summary": encrypt_if_present("The group is reviewing recursion before the midterm."),
+ "member_hash": "rich-member-hash-v1",
+ })
+ h.record("room_summaries", created=True)
+ else:
+ h.record("room_summaries", created=False)
+
+
 # (qa_id, concept_node_id, difficulty, score, total, questions_json, answers_json, completed_at)
 _QUIZ_ATTEMPTS = [
  ("rich-qa-cs-variables-1", "rich-node-cs-variables", "easy", 9, 10,
@@ -766,6 +782,7 @@ def seed_sessions() -> None:
  "schools", "courses", "course_offerings", "users", "user_profiles", "user_roles",
  "enrollments", "graph_nodes", "graph_edges", "node_mastery_events",
  "gradebook_categories", "assignments", "rooms", "room_members", "room_messages",
+ "room_summaries",
  "notes", "documents", "flashcards", "study_guides", "quiz_attempts", "quiz_context",
  "sessions", "messages", "feedback", "issue_reports",
 ]
@@ -785,6 +802,7 @@ def main() -> None:
  seed_notes_documents()
  seed_flashcards()
  seed_study_guides()
+ seed_room_summaries()
  seed_quiz()
  seed_feedback()
  seed_sessions()

backend/e2e_oracles/gather.py

@@ -195,6 +195,10 @@ def run_counts(args: argparse.Namespace) -> tuple[list[Finding], int]:
  ("quiz_attempts", "id", "questions_json"),
  ("quiz_attempts", "id", "answers_json"),
  ("quiz_context", "id", "context_json"),
+ ("flashcards", "id", "front"),
+ ("flashcards", "id", "back"),
+ ("study_guides", "id", "content"),
+ ("room_summaries", "room_id", "summary"),
 )
 
 

backend/routes/flashcards.py

@@ -14,7 +14,7 @@
 from services.academics import resolve_offering, term_id_for_label
 from services.auth_guard import require_self, get_session_user_id
 from services.achievement_service import check_achievements
-from services.encryption import decrypt_if_present, decrypt_json
+from services.encryption import decrypt_if_present, decrypt_json, encrypt_if_present
 from services.flashcard_import_service import (
  dedup_against_existing,
  check_rate_limit,
@@ -236,8 +236,8 @@ def generate(body: GenerateFlashcardsBody, request: Request):
  "id": str(uuid.uuid4()),
  "user_id": body.user_id,
  "topic": body.topic,
- "front": c["front"],
- "back": c["back"],
+ "front": encrypt_if_present(c["front"]),
+ "back": encrypt_if_present(c["back"]),
  "times_reviewed": 0,
  "last_reviewed_at": None,
  "created_at": now,
@@ -261,8 +261,16 @@ def generate(body: GenerateFlashcardsBody, request: Request):
  except Exception:
  pass
 
+ # rows_to_insert holds the ciphertext just written; the response the
+ # frontend renders (freshly generated cards, before any list re-fetch)
+ # must carry plaintext front/back, not the encrypted insert payload.
+ response_cards = [
+ {**row, "front": decrypt_if_present(row["front"]), "back": decrypt_if_present(row["back"])}
+ for row in rows_to_insert
+ ]
+
  return {
- "flashcards": rows_to_insert,
+ "flashcards": response_cards,
  "context_used": {
  "documents_found": len(documents),
  "weak_concepts_found": len(weak_concepts),
@@ -291,6 +299,9 @@ def get_flashcards(
  "id,user_id,topic,offering_id,front,back,times_reviewed,last_rating,last_reviewed_at,created_at",
  filters=filters, order="created_at.desc"
  ) or []
+ for r in rows:
+ r["front"] = decrypt_if_present(r.get("front"))
+ r["back"] = decrypt_if_present(r.get("back"))
  if semester:
  # Term scoping (#141). This route is user-wide (no course id), so
  # the filter works on the cards' offering: cards from the selected
@@ -397,8 +408,8 @@ def import_commit(body: ImportCommitBody, request: Request):
  "user_id": body.user_id,
  "topic": body.topic,
  "offering_id": offering_id,
- "front": c["front"],
- "back": c["back"],
+ "front": encrypt_if_present(c["front"]),
+ "back": encrypt_if_present(c["back"]),
  "times_reviewed": 0,
  "last_reviewed_at": None,
  "created_at": now,

backend/routes/study_guide.py

@@ -4,6 +4,7 @@
 Study guide generation and caching.
 """
 
+import logging
 import uuid
 from datetime import datetime, timezone
 
@@ -24,10 +25,17 @@
 )
 from services.graph_service import get_courses as graph_get_courses
 from services.auth_guard import require_self
-from services.encryption import decrypt_if_present, decrypt_json
+from services.encryption import (
+ decrypt_if_present,
+ decrypt_json,
+ decrypt_json_column,
+ encrypt_json,
+)
 from services.http_cache import cached_json, conditional, make_etag
 from services.request_context import current_request_id
 
+logger = logging.getLogger(__name__)
+
 router = APIRouter()
 
 
@@ -198,7 +206,7 @@ def _generate_and_insert(user_id: str, offering_id: str, exam_id: str) -> dict:
  "offering_id": offering_id,
  "exam_id": exam_id,
  "generated_at": now,
- "content": content,
+ "content": encrypt_json(content),
  }
  table("study_guides").insert(row)
 
@@ -245,7 +253,14 @@ def get_cached_guides(user_id: str, request: Request):
 
  result = []
  for g in guides:
- content = g.get("content") or {}
+ try:
+ content = decrypt_json_column(g.get("content")) or {}
+ except Exception:
+ logger.warning(
+ "get_cached_guides: content decrypt failed for guide %s; degrading",
+ g.get("id"),
+ )
+ content = {}
  course_id = offering_to_course.get(g.get("offering_id"))
  result.append({
  "id": g["id"],
@@ -330,7 +345,7 @@ def get_guide(
  )
  if cached:
  row = cached[0]
- return {"guide": row["content"], "generated_at": row["generated_at"], "cached": True}
+ return {"guide": decrypt_json_column(row["content"]), "generated_at": row["generated_at"], "cached": True}
 
  result = _generate_and_insert(user_id, offering_id, exam_id)
  return {"guide": result["content"], "generated_at": result["generated_at"], "cached": False}

backend/services/flashcard_import_service.py

@@ -33,6 +33,7 @@
 from agents.flashcard import flashcard_agent
 from agents.usage import record_agent_usage
 from services import extraction_service
+from services.encryption import decrypt_if_present
 
 logger = logging.getLogger(__name__)
 
@@ -76,7 +77,9 @@ def dedup_against_existing(
  filters["topic"] = f"eq.{topic}"
 
  existing = table("flashcards").select("front", filters=filters) or []
- existing_norm = [_normalize(r.get("front", "")) for r in existing]
+ existing_norm = [
+ _normalize(decrypt_if_present(r.get("front", "")) or "") for r in existing
+ ]
 
  keep: list[Card] = []
  skipped: list[Card] = []

backend/services/social_cache_service.py

@@ -12,6 +12,7 @@
 from datetime import datetime, timezone
 
 from db.connection import table
+from services.encryption import decrypt_if_present, encrypt_if_present
 
 
 def _compute_hash(member_summaries: list[str]) -> str:
@@ -26,15 +27,15 @@ def get_cached_summary(room_id: str, member_summaries: list[str]) -> str | None:
  filters={"room_id": f"eq.{room_id}"},
  )
  if rows and rows[0]["member_hash"] == current_hash:
- return rows[0]["summary"]
+ return decrypt_if_present(rows[0]["summary"])
  return None
 
 
 def save_summary(room_id: str, member_summaries: list[str], summary: str) -> None:
  table("room_summaries").upsert(
  {
  "room_id": room_id,
- "summary": summary,
+ "summary": encrypt_if_present(summary),
  "member_hash": _compute_hash(member_summaries),
  "updated_at": datetime.now(timezone.utc).isoformat(),
  },

backend/tests/integration/test_encryption_roundtrip.py

@@ -46,6 +46,13 @@
  "Upload stuck"),
  ("issue_reports.description", "issue_reports", "id", "rich-issue-1",
  "description", "Syllabus upload spins forever."),
+ ("flashcards.front", "flashcards", "id", "rich-fc-cs-1", "front",
+ "What is a variable?"),
+ ("flashcards.back", "flashcards", "id", "rich-fc-cs-1", "back",
+ "A named storage location for a value."),
+ ("room_summaries.summary", "room_summaries", "room_id",
+ "rich-room-study-group", "summary",
+ "The group is reviewing recursion before the midterm."),
 ]
 
 # (label, id_value, column, expected_number)
@@ -123,6 +130,14 @@ def test_quiz_context_context_json_is_ciphertext_and_decrypts(db_conn):
  assert decoded["asked"] == 2
 
 
+def test_study_guides_content_is_ciphertext_and_decrypts(db_conn):
+ """#518: study_guides.content needs the JSON pair (encrypt_json/decrypt_json)."""
+ raw = _raw(db_conn, "study_guides", "id", "rich-guide-cs-f25-mid", "content")
+ assert isinstance(raw, str), "content stored as PLAINTEXT JSONB — encryption regressed"
+ decoded = decrypt_json(raw)
+ assert decoded["exam"] == "Midterm Exam"
+
+
 def test_documents_concept_notes_uses_json_encryption(db_conn):
  """documents.concept_notes is stored via encrypt_json (a list of concepts)."""
  raw = _raw(db_conn, "documents", "id", "rich-doc-cs-syllabus", "concept_notes")

backend/tests/test_flashcard_import_service.py

@@ -59,6 +59,28 @@ def test_filters_by_topic_when_offering_id_is_none(self):
  assert "offering_id" not in str(call_kwargs)
  assert "course_id" not in str(call_kwargs)
 
+ def test_dedup_matches_against_encrypted_existing_cards(self):
+ # #518: dedupe must be decrypt-aware — existing rows are ciphertext.
+ from services.encryption import encrypt
+ with patch("services.flashcard_import_service.table") as t:
+ t.return_value.select.return_value = [
+ {"front": encrypt("What is a variable?")}
+ ]
+ new = [{"front": "What is a variable?", "back": "A named storage location."}]
+ keep, skipped = svc.dedup_against_existing("u1", "c1", new)
+ assert keep == []
+ assert skipped == new
+
+ def test_dedup_still_matches_legacy_plaintext_rows(self):
+ # Pre-backfill rows are still plaintext — decrypt_if_present falls
+ # back to the raw value, so dedupe keeps working unchanged.
+ with patch("services.flashcard_import_service.table") as t:
+ t.return_value.select.return_value = [{"front": "What is a variable?"}]
+ new = [{"front": "What is a variable?", "back": "A named storage location."}]
+ keep, skipped = svc.dedup_against_existing("u1", "c1", new)
+ assert keep == []
+ assert skipped == new
+
 
 # ── check_rate_limit ─────────────────────────────────────────────────────────
 

CLAUDE.md

@@ -106,7 +106,7 @@ make explore # Chapter 2: bounded AI exploration of the ru
 
 ## Gotchas
 
-- Column-level encryption is on for sensitive columns: `user_profiles.name`/`first_name`/`last_name`/`bio`/`location` (these moved off `users` to `user_profiles` in the 0024 identity split), Google OAuth tokens, `messages.content`, `room_messages.text`, `sessions.summary_json`, `documents.summary` + `concept_notes` + `extracted_text` (the RAG OCR text added in 0030), `notes.title`/`body`/`last_summary`, `assignments.notes`/`points_possible`/`points_earned` (the enrollment-keyed gradebook table; points columns carry numeric semantics — use `decrypt_numeric` at read), `feedback.comment`/`topic` + `issue_reports.topic`/`description` (free-text user input, #520), and `quiz_attempts.questions_json`/`answers_json` + `quiz_context.context_json` (quiz performance data, #521; scalar analytics columns — score/total/difficulty/completed_at — stay plaintext). Helpers live in `backend/services/encryption.py`; use `encrypt_if_present` at write boundaries and `decrypt_if_present` / `decrypt_numeric` at read boundaries (including before injecting into AI prompts). `ENCRYPTION_KEY` must be set (32 bytes as 64 hex chars; generate via `python -c "import secrets; print(secrets.token_hex(32))"`). Deliberate exception: `newsletter_emails.email` stays plaintext (ADR 0026) — the UNIQUE constraint, lookup index, and both subscribe/allowlist upserts key on the value, and AES-GCM's per-call nonce breaks value equality.
+- Column-level encryption is on for sensitive columns: `user_profiles.name`/`first_name`/`last_name`/`bio`/`location` (these moved off `users` to `user_profiles` in the 0024 identity split), Google OAuth tokens, `messages.content`, `room_messages.text`, `sessions.summary_json`, `documents.summary` + `concept_notes` + `extracted_text` (the RAG OCR text added in 0030), `notes.title`/`body`/`last_summary`, `assignments.notes`/`points_possible`/`points_earned` (the enrollment-keyed gradebook table; points columns carry numeric semantics — use `decrypt_numeric` at read), `feedback.comment`/`topic` + `issue_reports.topic`/`description` (free-text user input, #520), `quiz_attempts.questions_json`/`answers_json` + `quiz_context.context_json` (quiz performance data, #521; scalar analytics columns — score/total/difficulty/completed_at — stay plaintext), and `flashcards.front`/`back` + `study_guides.content` + `room_summaries.summary` (derived content, #518; `study_guides.content` uses the JSON pair, `room_summaries.summary` keys its cache on the separate plaintext `member_hash` column so encryption doesn't affect cache-hit lookups). Helpers live in `backend/services/encryption.py`; use `encrypt_if_present` at write boundaries and `decrypt_if_present` / `decrypt_numeric` at read boundaries (including before injecting into AI prompts). `ENCRYPTION_KEY` must be set (32 bytes as 64 hex chars; generate via `python -c "import secrets; print(secrets.token_hex(32))"`). Deliberate exception: `newsletter_emails.email` stays plaintext (ADR 0026) — the UNIQUE constraint, lookup index, and both subscribe/allowlist upserts key on the value, and AES-GCM's per-call nonce breaks value equality.
 - Knowledge-graph mastery is now an append-only `node_mastery_events` table (replaced the `graph_nodes.mastery_events` JSON column in 0023); node/edge dedup is enforced by UNIQUE constraints. Don't read/write a `mastery_events` column.
 - Optional cross-worker cache (#97): `services/cache.py` wraps Redis and is **off by default** — with no `REDIS_URL` set it's a zero-overhead no-op and never fails a request (any Redis error → clean miss + warning). Currently backs the content-addressed OCR/extraction cache (`extraction_service.extract_text_from_file`, keyed on `sha256(file_bytes)` + engine). The `redis` dependency is only imported when `REDIS_URL` is set.
 - HTTP caching (#99): conditional GETs use `services/http_cache.py` (`make_etag`/`conditional`/`cached_json`). `Cache-Control` on these routes is **always `private`, never `public`** — the responses carry user-scoped, app-decrypted columns that must never be cached at a shared proxy/CDN. Derive the ETag from cheap change-keys (ids, `updated_at`, existing content hashes), not from the fully-built payload.

backend/db/backfill_encryption.py

@@ -302,6 +302,18 @@ def backfill_quiz_context(apply: bool) -> dict:
  return _encrypt_json_column("quiz_context", "context_json", pk="id", apply=apply)
 
 
+def backfill_flashcards(apply: bool) -> dict:
+ return _encrypt_text_column("flashcards", ["front", "back"], pk="id", apply=apply)
+
+
+def backfill_study_guides(apply: bool) -> dict:
+ return _encrypt_json_column("study_guides", "content", pk="id", apply=apply)
+
+
+def backfill_room_summaries(apply: bool) -> dict:
+ return _encrypt_text_column("room_summaries", ["summary"], pk="room_id", apply=apply)
+
+
 RUNNERS: dict[str, Callable[[bool], dict]] = {
  "users": backfill_users,
  "user_settings": backfill_user_settings,
@@ -315,6 +327,9 @@ def backfill_quiz_context(apply: bool) -> dict:
  "issue_reports": backfill_issue_reports,
  "quiz_attempts": backfill_quiz_attempts,
  "quiz_context": backfill_quiz_context,
+ "flashcards": backfill_flashcards,
+ "study_guides": backfill_study_guides,
+ "room_summaries": backfill_room_summaries,
 }
 
 

backend/db/seed_local_rich.py

@@ -553,7 +553,7 @@ def seed_notes_documents() -> None:
  )
 
 
-# (fc_id, user_id, offering_id, topic, front, back) — plaintext, grouped by topic.
+# (fc_id, user_id, offering_id, topic, front, back) — 🔒 front/back, grouped by topic.
 _FLASHCARDS = [
  ("rich-fc-cs-1", USER_ACTIVE, OFF_CS_F25, "CS Basics",
  "What is a variable?", "A named storage location for a value."),
@@ -579,8 +579,9 @@ def seed_flashcards() -> None:
  "user_id": user_id,
  "offering_id": off_id,
  "topic": topic,
- "front": front,
- "back": back,
+ # 🔒 front / back (#518)
+ "front": encrypt_if_present(front),
+ "back": encrypt_if_present(back),
  },
  )
 
@@ -625,11 +626,26 @@ def seed_study_guides() -> None:
  "offering_id": off_id,
  "exam_id": exam_id,
  "generated_at": generated_at,
- "content": content,
+ # 🔒 content (#518)
+ "content": encrypt_json(content),
  },
  )
 
 
+def seed_room_summaries() -> None:
+ # #518: room_summaries.summary is 🔒. PK is room_id (no id column), so this
+ # can't go through insert_if_absent.
+ if not table("room_summaries").select("room_id", filters={"room_id": f"eq.{ROOM_STUDY}"}):
+ table("room_summaries").insert({
+ "room_id": ROOM_STUDY,
+ "summary": encrypt_if_present("The group is reviewing recursion before the midterm."),
+ "member_hash": "rich-member-hash-v1",
+ })
+ h.record("room_summaries", created=True)
+ else:
+ h.record("room_summaries", created=False)
+
+
 # (qa_id, concept_node_id, difficulty, score, total, questions_json, answers_json, completed_at)
 _QUIZ_ATTEMPTS = [
  ("rich-qa-cs-variables-1", "rich-node-cs-variables", "easy", 9, 10,
@@ -766,6 +782,7 @@ def seed_sessions() -> None:
  "schools", "courses", "course_offerings", "users", "user_profiles", "user_roles",
  "enrollments", "graph_nodes", "graph_edges", "node_mastery_events",
  "gradebook_categories", "assignments", "rooms", "room_members", "room_messages",
+ "room_summaries",
  "notes", "documents", "flashcards", "study_guides", "quiz_attempts", "quiz_context",
  "sessions", "messages", "feedback", "issue_reports",
 ]
@@ -785,6 +802,7 @@ def main() -> None:
  seed_notes_documents()
  seed_flashcards()
  seed_study_guides()
+ seed_room_summaries()
  seed_quiz()
  seed_feedback()
  seed_sessions()

backend/e2e_oracles/gather.py

@@ -195,6 +195,10 @@ def run_counts(args: argparse.Namespace) -> tuple[list[Finding], int]:
  ("quiz_attempts", "id", "questions_json"),
  ("quiz_attempts", "id", "answers_json"),
  ("quiz_context", "id", "context_json"),
+ ("flashcards", "id", "front"),
+ ("flashcards", "id", "back"),
+ ("study_guides", "id", "content"),
+ ("room_summaries", "room_id", "summary"),
 )
 
 

backend/routes/flashcards.py

@@ -14,7 +14,7 @@
 from services.academics import resolve_offering, term_id_for_label
 from services.auth_guard import require_self, get_session_user_id
 from services.achievement_service import check_achievements
-from services.encryption import decrypt_if_present, decrypt_json
+from services.encryption import decrypt_if_present, decrypt_json, encrypt_if_present
 from services.flashcard_import_service import (
  dedup_against_existing,
  check_rate_limit,
@@ -236,8 +236,8 @@ def generate(body: GenerateFlashcardsBody, request: Request):
  "id": str(uuid.uuid4()),
  "user_id": body.user_id,
  "topic": body.topic,
- "front": c["front"],
- "back": c["back"],
+ "front": encrypt_if_present(c["front"]),
+ "back": encrypt_if_present(c["back"]),
  "times_reviewed": 0,
  "last_reviewed_at": None,
  "created_at": now,
@@ -261,8 +261,16 @@ def generate(body: GenerateFlashcardsBody, request: Request):
  except Exception:
  pass
 
+ # rows_to_insert holds the ciphertext just written; the response the
+ # frontend renders (freshly generated cards, before any list re-fetch)
+ # must carry plaintext front/back, not the encrypted insert payload.
+ response_cards = [
+ {**row, "front": decrypt_if_present(row["front"]), "back": decrypt_if_present(row["back"])}
+ for row in rows_to_insert
+ ]
+
  return {
- "flashcards": rows_to_insert,
+ "flashcards": response_cards,
  "context_used": {
  "documents_found": len(documents),
  "weak_concepts_found": len(weak_concepts),
@@ -291,6 +299,9 @@ def get_flashcards(
  "id,user_id,topic,offering_id,front,back,times_reviewed,last_rating,last_reviewed_at,created_at",
  filters=filters, order="created_at.desc"
  ) or []
+ for r in rows:
+ r["front"] = decrypt_if_present(r.get("front"))
+ r["back"] = decrypt_if_present(r.get("back"))
  if semester:
  # Term scoping (#141). This route is user-wide (no course id), so
  # the filter works on the cards' offering: cards from the selected
@@ -397,8 +408,8 @@ def import_commit(body: ImportCommitBody, request: Request):
  "user_id": body.user_id,
  "topic": body.topic,
  "offering_id": offering_id,
- "front": c["front"],
- "back": c["back"],
+ "front": encrypt_if_present(c["front"]),
+ "back": encrypt_if_present(c["back"]),
  "times_reviewed": 0,
  "last_reviewed_at": None,
  "created_at": now,

backend/routes/study_guide.py

@@ -4,6 +4,7 @@
 Study guide generation and caching.
 """
 
+import logging
 import uuid
 from datetime import datetime, timezone
 
@@ -24,10 +25,17 @@
 )
 from services.graph_service import get_courses as graph_get_courses
 from services.auth_guard import require_self
-from services.encryption import decrypt_if_present, decrypt_json
+from services.encryption import (
+ decrypt_if_present,
+ decrypt_json,
+ decrypt_json_column,
+ encrypt_json,
+)
 from services.http_cache import cached_json, conditional, make_etag
 from services.request_context import current_request_id
 
+logger = logging.getLogger(__name__)
+
 router = APIRouter()
 
 
@@ -198,7 +206,7 @@ def _generate_and_insert(user_id: str, offering_id: str, exam_id: str) -> dict:
  "offering_id": offering_id,
  "exam_id": exam_id,
  "generated_at": now,
- "content": content,
+ "content": encrypt_json(content),
  }
  table("study_guides").insert(row)
 
@@ -245,7 +253,14 @@ def get_cached_guides(user_id: str, request: Request):
 
  result = []
  for g in guides:
- content = g.get("content") or {}
+ try:
+ content = decrypt_json_column(g.get("content")) or {}
+ except Exception:
+ logger.warning(
+ "get_cached_guides: content decrypt failed for guide %s; degrading",
+ g.get("id"),
+ )
+ content = {}
  course_id = offering_to_course.get(g.get("offering_id"))
  result.append({
  "id": g["id"],
@@ -330,7 +345,7 @@ def get_guide(
  )
  if cached:
  row = cached[0]
- return {"guide": row["content"], "generated_at": row["generated_at"], "cached": True}
+ return {"guide": decrypt_json_column(row["content"]), "generated_at": row["generated_at"], "cached": True}
 
  result = _generate_and_insert(user_id, offering_id, exam_id)
  return {"guide": result["content"], "generated_at": result["generated_at"], "cached": False}

backend/services/flashcard_import_service.py

@@ -33,6 +33,7 @@
 from agents.flashcard import flashcard_agent
 from agents.usage import record_agent_usage
 from services import extraction_service
+from services.encryption import decrypt_if_present
 
 logger = logging.getLogger(__name__)
 
@@ -76,7 +77,9 @@ def dedup_against_existing(
  filters["topic"] = f"eq.{topic}"
 
  existing = table("flashcards").select("front", filters=filters) or []
- existing_norm = [_normalize(r.get("front", "")) for r in existing]
+ existing_norm = [
+ _normalize(decrypt_if_present(r.get("front", "")) or "") for r in existing
+ ]
 
  keep: list[Card] = []
  skipped: list[Card] = []

backend/services/social_cache_service.py

@@ -12,6 +12,7 @@
 from datetime import datetime, timezone
 
 from db.connection import table
+from services.encryption import decrypt_if_present, encrypt_if_present
 
 
 def _compute_hash(member_summaries: list[str]) -> str:
@@ -26,15 +27,15 @@ def get_cached_summary(room_id: str, member_summaries: list[str]) -> str | None:
  filters={"room_id": f"eq.{room_id}"},
  )
  if rows and rows[0]["member_hash"] == current_hash:
- return rows[0]["summary"]
+ return decrypt_if_present(rows[0]["summary"])
  return None
 
 
 def save_summary(room_id: str, member_summaries: list[str], summary: str) -> None:
  table("room_summaries").upsert(
  {
  "room_id": room_id,
- "summary": summary,
+ "summary": encrypt_if_present(summary),
  "member_hash": _compute_hash(member_summaries),
  "updated_at": datetime.now(timezone.utc).isoformat(),
  },

backend/tests/integration/test_encryption_roundtrip.py

@@ -46,6 +46,13 @@
  "Upload stuck"),
  ("issue_reports.description", "issue_reports", "id", "rich-issue-1",
  "description", "Syllabus upload spins forever."),
+ ("flashcards.front", "flashcards", "id", "rich-fc-cs-1", "front",
+ "What is a variable?"),
+ ("flashcards.back", "flashcards", "id", "rich-fc-cs-1", "back",
+ "A named storage location for a value."),
+ ("room_summaries.summary", "room_summaries", "room_id",
+ "rich-room-study-group", "summary",
+ "The group is reviewing recursion before the midterm."),
 ]
 
 # (label, id_value, column, expected_number)
@@ -123,6 +130,14 @@ def test_quiz_context_context_json_is_ciphertext_and_decrypts(db_conn):
  assert decoded["asked"] == 2
 
 
+def test_study_guides_content_is_ciphertext_and_decrypts(db_conn):
+ """#518: study_guides.content needs the JSON pair (encrypt_json/decrypt_json)."""
+ raw = _raw(db_conn, "study_guides", "id", "rich-guide-cs-f25-mid", "content")
+ assert isinstance(raw, str), "content stored as PLAINTEXT JSONB — encryption regressed"
+ decoded = decrypt_json(raw)
+ assert decoded["exam"] == "Midterm Exam"
+
+
 def test_documents_concept_notes_uses_json_encryption(db_conn):
  """documents.concept_notes is stored via encrypt_json (a list of concepts)."""
  raw = _raw(db_conn, "documents", "id", "rich-doc-cs-syllabus", "concept_notes")

backend/tests/test_flashcard_import_service.py

@@ -59,6 +59,28 @@ def test_filters_by_topic_when_offering_id_is_none(self):
  assert "offering_id" not in str(call_kwargs)
  assert "course_id" not in str(call_kwargs)
 
+ def test_dedup_matches_against_encrypted_existing_cards(self):
+ # #518: dedupe must be decrypt-aware — existing rows are ciphertext.
+ from services.encryption import encrypt
+ with patch("services.flashcard_import_service.table") as t:
+ t.return_value.select.return_value = [
+ {"front": encrypt("What is a variable?")}
+ ]
+ new = [{"front": "What is a variable?", "back": "A named storage location."}]
+ keep, skipped = svc.dedup_against_existing("u1", "c1", new)
+ assert keep == []
+ assert skipped == new
+
+ def test_dedup_still_matches_legacy_plaintext_rows(self):
+ # Pre-backfill rows are still plaintext — decrypt_if_present falls
+ # back to the raw value, so dedupe keeps working unchanged.
+ with patch("services.flashcard_import_service.table") as t:
+ t.return_value.select.return_value = [{"front": "What is a variable?"}]
+ new = [{"front": "What is a variable?", "back": "A named storage location."}]
+ keep, skipped = svc.dedup_against_existing("u1", "c1", new)
+ assert keep == []
+ assert skipped == new
+
 
 # ── check_rate_limit ─────────────────────────────────────────────────────────
 

CLAUDE.md

@@ -106,7 +106,7 @@ make explore # Chapter 2: bounded AI exploration of the ru
 
 ## Gotchas
 
-- Column-level encryption is on for sensitive columns: `user_profiles.name`/`first_name`/`last_name`/`bio`/`location` (these moved off `users` to `user_profiles` in the 0024 identity split), Google OAuth tokens, `messages.content`, `room_messages.text`, `sessions.summary_json`, `documents.summary` + `concept_notes` + `extracted_text` (the RAG OCR text added in 0030), `notes.title`/`body`/`last_summary`, `assignments.notes`/`points_possible`/`points_earned` (the enrollment-keyed gradebook table; points columns carry numeric semantics — use `decrypt_numeric` at read), `feedback.comment`/`topic` + `issue_reports.topic`/`description` (free-text user input, #520), and `quiz_attempts.questions_json`/`answers_json` + `quiz_context.context_json` (quiz performance data, #521; scalar analytics columns — score/total/difficulty/completed_at — stay plaintext). Helpers live in `backend/services/encryption.py`; use `encrypt_if_present` at write boundaries and `decrypt_if_present` / `decrypt_numeric` at read boundaries (including before injecting into AI prompts). `ENCRYPTION_KEY` must be set (32 bytes as 64 hex chars; generate via `python -c "import secrets; print(secrets.token_hex(32))"`). Deliberate exception: `newsletter_emails.email` stays plaintext (ADR 0026) — the UNIQUE constraint, lookup index, and both subscribe/allowlist upserts key on the value, and AES-GCM's per-call nonce breaks value equality.
+- Column-level encryption is on for sensitive columns: `user_profiles.name`/`first_name`/`last_name`/`bio`/`location` (these moved off `users` to `user_profiles` in the 0024 identity split), Google OAuth tokens, `messages.content`, `room_messages.text`, `sessions.summary_json`, `documents.summary` + `concept_notes` + `extracted_text` (the RAG OCR text added in 0030), `notes.title`/`body`/`last_summary`, `assignments.notes`/`points_possible`/`points_earned` (the enrollment-keyed gradebook table; points columns carry numeric semantics — use `decrypt_numeric` at read), `feedback.comment`/`topic` + `issue_reports.topic`/`description` (free-text user input, #520), `quiz_attempts.questions_json`/`answers_json` + `quiz_context.context_json` (quiz performance data, #521; scalar analytics columns — score/total/difficulty/completed_at — stay plaintext), and `flashcards.front`/`back` + `study_guides.content` + `room_summaries.summary` (derived content, #518; `study_guides.content` uses the JSON pair, `room_summaries.summary` keys its cache on the separate plaintext `member_hash` column so encryption doesn't affect cache-hit lookups). Helpers live in `backend/services/encryption.py`; use `encrypt_if_present` at write boundaries and `decrypt_if_present` / `decrypt_numeric` at read boundaries (including before injecting into AI prompts). `ENCRYPTION_KEY` must be set (32 bytes as 64 hex chars; generate via `python -c "import secrets; print(secrets.token_hex(32))"`). Deliberate exception: `newsletter_emails.email` stays plaintext (ADR 0026) — the UNIQUE constraint, lookup index, and both subscribe/allowlist upserts key on the value, and AES-GCM's per-call nonce breaks value equality.
 - Knowledge-graph mastery is now an append-only `node_mastery_events` table (replaced the `graph_nodes.mastery_events` JSON column in 0023); node/edge dedup is enforced by UNIQUE constraints. Don't read/write a `mastery_events` column.
 - Optional cross-worker cache (#97): `services/cache.py` wraps Redis and is **off by default** — with no `REDIS_URL` set it's a zero-overhead no-op and never fails a request (any Redis error → clean miss + warning). Currently backs the content-addressed OCR/extraction cache (`extraction_service.extract_text_from_file`, keyed on `sha256(file_bytes)` + engine). The `redis` dependency is only imported when `REDIS_URL` is set.
 - HTTP caching (#99): conditional GETs use `services/http_cache.py` (`make_etag`/`conditional`/`cached_json`). `Cache-Control` on these routes is **always `private`, never `public`** — the responses carry user-scoped, app-decrypted columns that must never be cached at a shared proxy/CDN. Derive the ETag from cheap change-keys (ids, `updated_at`, existing content hashes), not from the fully-built payload.

backend/db/backfill_encryption.py

@@ -302,6 +302,18 @@ def backfill_quiz_context(apply: bool) -> dict:
  return _encrypt_json_column("quiz_context", "context_json", pk="id", apply=apply)
 
 
+def backfill_flashcards(apply: bool) -> dict:
+ return _encrypt_text_column("flashcards", ["front", "back"], pk="id", apply=apply)
+
+
+def backfill_study_guides(apply: bool) -> dict:
+ return _encrypt_json_column("study_guides", "content", pk="id", apply=apply)
+
+
+def backfill_room_summaries(apply: bool) -> dict:
+ return _encrypt_text_column("room_summaries", ["summary"], pk="room_id", apply=apply)
+
+
 RUNNERS: dict[str, Callable[[bool], dict]] = {
  "users": backfill_users,
  "user_settings": backfill_user_settings,
@@ -315,6 +327,9 @@ def backfill_quiz_context(apply: bool) -> dict:
  "issue_reports": backfill_issue_reports,
  "quiz_attempts": backfill_quiz_attempts,
  "quiz_context": backfill_quiz_context,
+ "flashcards": backfill_flashcards,
+ "study_guides": backfill_study_guides,
+ "room_summaries": backfill_room_summaries,
 }
 
 

backend/db/seed_local_rich.py

@@ -553,7 +553,7 @@ def seed_notes_documents() -> None:
  )
 
 
-# (fc_id, user_id, offering_id, topic, front, back) — plaintext, grouped by topic.
+# (fc_id, user_id, offering_id, topic, front, back) — 🔒 front/back, grouped by topic.
 _FLASHCARDS = [
  ("rich-fc-cs-1", USER_ACTIVE, OFF_CS_F25, "CS Basics",
  "What is a variable?", "A named storage location for a value."),
@@ -579,8 +579,9 @@ def seed_flashcards() -> None:
  "user_id": user_id,
  "offering_id": off_id,
  "topic": topic,
- "front": front,
- "back": back,
+ # 🔒 front / back (#518)
+ "front": encrypt_if_present(front),
+ "back": encrypt_if_present(back),
  },
  )
 
@@ -625,11 +626,26 @@ def seed_study_guides() -> None:
  "offering_id": off_id,
  "exam_id": exam_id,
  "generated_at": generated_at,
- "content": content,
+ # 🔒 content (#518)
+ "content": encrypt_json(content),
  },
  )
 
 
+def seed_room_summaries() -> None:
+ # #518: room_summaries.summary is 🔒. PK is room_id (no id column), so this
+ # can't go through insert_if_absent.
+ if not table("room_summaries").select("room_id", filters={"room_id": f"eq.{ROOM_STUDY}"}):
+ table("room_summaries").insert({
+ "room_id": ROOM_STUDY,
+ "summary": encrypt_if_present("The group is reviewing recursion before the midterm."),
+ "member_hash": "rich-member-hash-v1",
+ })
+ h.record("room_summaries", created=True)
+ else:
+ h.record("room_summaries", created=False)
+
+
 # (qa_id, concept_node_id, difficulty, score, total, questions_json, answers_json, completed_at)
 _QUIZ_ATTEMPTS = [
  ("rich-qa-cs-variables-1", "rich-node-cs-variables", "easy", 9, 10,
@@ -766,6 +782,7 @@ def seed_sessions() -> None:
  "schools", "courses", "course_offerings", "users", "user_profiles", "user_roles",
  "enrollments", "graph_nodes", "graph_edges", "node_mastery_events",
  "gradebook_categories", "assignments", "rooms", "room_members", "room_messages",
+ "room_summaries",
  "notes", "documents", "flashcards", "study_guides", "quiz_attempts", "quiz_context",
  "sessions", "messages", "feedback", "issue_reports",
 ]
@@ -785,6 +802,7 @@ def main() -> None:
  seed_notes_documents()
  seed_flashcards()
  seed_study_guides()
+ seed_room_summaries()
  seed_quiz()
  seed_feedback()
  seed_sessions()

backend/e2e_oracles/gather.py

@@ -195,6 +195,10 @@ def run_counts(args: argparse.Namespace) -> tuple[list[Finding], int]:
  ("quiz_attempts", "id", "questions_json"),
  ("quiz_attempts", "id", "answers_json"),
  ("quiz_context", "id", "context_json"),
+ ("flashcards", "id", "front"),
+ ("flashcards", "id", "back"),
+ ("study_guides", "id", "content"),
+ ("room_summaries", "room_id", "summary"),
 )
 
 

backend/routes/flashcards.py

@@ -14,7 +14,7 @@
 from services.academics import resolve_offering, term_id_for_label
 from services.auth_guard import require_self, get_session_user_id
 from services.achievement_service import check_achievements
-from services.encryption import decrypt_if_present, decrypt_json
+from services.encryption import decrypt_if_present, decrypt_json, encrypt_if_present
 from services.flashcard_import_service import (
  dedup_against_existing,
  check_rate_limit,
@@ -236,8 +236,8 @@ def generate(body: GenerateFlashcardsBody, request: Request):
  "id": str(uuid.uuid4()),
  "user_id": body.user_id,
  "topic": body.topic,
- "front": c["front"],
- "back": c["back"],
+ "front": encrypt_if_present(c["front"]),
+ "back": encrypt_if_present(c["back"]),
  "times_reviewed": 0,
  "last_reviewed_at": None,
  "created_at": now,
@@ -261,8 +261,16 @@ def generate(body: GenerateFlashcardsBody, request: Request):
  except Exception:
  pass
 
+ # rows_to_insert holds the ciphertext just written; the response the
+ # frontend renders (freshly generated cards, before any list re-fetch)
+ # must carry plaintext front/back, not the encrypted insert payload.
+ response_cards = [
+ {**row, "front": decrypt_if_present(row["front"]), "back": decrypt_if_present(row["back"])}
+ for row in rows_to_insert
+ ]
+
  return {
- "flashcards": rows_to_insert,
+ "flashcards": response_cards,
  "context_used": {
  "documents_found": len(documents),
  "weak_concepts_found": len(weak_concepts),
@@ -291,6 +299,9 @@ def get_flashcards(
  "id,user_id,topic,offering_id,front,back,times_reviewed,last_rating,last_reviewed_at,created_at",
  filters=filters, order="created_at.desc"
  ) or []
+ for r in rows:
+ r["front"] = decrypt_if_present(r.get("front"))
+ r["back"] = decrypt_if_present(r.get("back"))
  if semester:
  # Term scoping (#141). This route is user-wide (no course id), so
  # the filter works on the cards' offering: cards from the selected
@@ -397,8 +408,8 @@ def import_commit(body: ImportCommitBody, request: Request):
  "user_id": body.user_id,
  "topic": body.topic,
  "offering_id": offering_id,
- "front": c["front"],
- "back": c["back"],
+ "front": encrypt_if_present(c["front"]),
+ "back": encrypt_if_present(c["back"]),
  "times_reviewed": 0,
  "last_reviewed_at": None,
  "created_at": now,

backend/routes/study_guide.py

@@ -4,6 +4,7 @@
 Study guide generation and caching.
 """
 
+import logging
 import uuid
 from datetime import datetime, timezone
 
@@ -24,10 +25,17 @@
 )
 from services.graph_service import get_courses as graph_get_courses
 from services.auth_guard import require_self
-from services.encryption import decrypt_if_present, decrypt_json
+from services.encryption import (
+ decrypt_if_present,
+ decrypt_json,
+ decrypt_json_column,
+ encrypt_json,
+)
 from services.http_cache import cached_json, conditional, make_etag
 from services.request_context import current_request_id
 
+logger = logging.getLogger(__name__)
+
 router = APIRouter()
 
 
@@ -198,7 +206,7 @@ def _generate_and_insert(user_id: str, offering_id: str, exam_id: str) -> dict:
  "offering_id": offering_id,
  "exam_id": exam_id,
  "generated_at": now,
- "content": content,
+ "content": encrypt_json(content),
  }
  table("study_guides").insert(row)
 
@@ -245,7 +253,14 @@ def get_cached_guides(user_id: str, request: Request):
 
  result = []
  for g in guides:
- content = g.get("content") or {}
+ try:
+ content = decrypt_json_column(g.get("content")) or {}
+ except Exception:
+ logger.warning(
+ "get_cached_guides: content decrypt failed for guide %s; degrading",
+ g.get("id"),
+ )
+ content = {}
  course_id = offering_to_course.get(g.get("offering_id"))
  result.append({
  "id": g["id"],
@@ -330,7 +345,7 @@ def get_guide(
  )
  if cached:
  row = cached[0]
- return {"guide": row["content"], "generated_at": row["generated_at"], "cached": True}
+ return {"guide": decrypt_json_column(row["content"]), "generated_at": row["generated_at"], "cached": True}
 
  result = _generate_and_insert(user_id, offering_id, exam_id)
  return {"guide": result["content"], "generated_at": result["generated_at"], "cached": False}

backend/services/flashcard_import_service.py

@@ -33,6 +33,7 @@
 from agents.flashcard import flashcard_agent
 from agents.usage import record_agent_usage
 from services import extraction_service
+from services.encryption import decrypt_if_present
 
 logger = logging.getLogger(__name__)
 
@@ -76,7 +77,9 @@ def dedup_against_existing(
  filters["topic"] = f"eq.{topic}"
 
  existing = table("flashcards").select("front", filters=filters) or []
- existing_norm = [_normalize(r.get("front", "")) for r in existing]
+ existing_norm = [
+ _normalize(decrypt_if_present(r.get("front", "")) or "") for r in existing
+ ]
 
  keep: list[Card] = []
  skipped: list[Card] = []

backend/services/social_cache_service.py

@@ -12,6 +12,7 @@
 from datetime import datetime, timezone
 
 from db.connection import table
+from services.encryption import decrypt_if_present, encrypt_if_present
 
 
 def _compute_hash(member_summaries: list[str]) -> str:
@@ -26,15 +27,15 @@ def get_cached_summary(room_id: str, member_summaries: list[str]) -> str | None:
  filters={"room_id": f"eq.{room_id}"},
  )
  if rows and rows[0]["member_hash"] == current_hash:
- return rows[0]["summary"]
+ return decrypt_if_present(rows[0]["summary"])
  return None
 
 
 def save_summary(room_id: str, member_summaries: list[str], summary: str) -> None:
  table("room_summaries").upsert(
  {
  "room_id": room_id,
- "summary": summary,
+ "summary": encrypt_if_present(summary),
  "member_hash": _compute_hash(member_summaries),
  "updated_at": datetime.now(timezone.utc).isoformat(),
  },

backend/tests/integration/test_encryption_roundtrip.py

@@ -46,6 +46,13 @@
  "Upload stuck"),
  ("issue_reports.description", "issue_reports", "id", "rich-issue-1",
  "description", "Syllabus upload spins forever."),
+ ("flashcards.front", "flashcards", "id", "rich-fc-cs-1", "front",
+ "What is a variable?"),
+ ("flashcards.back", "flashcards", "id", "rich-fc-cs-1", "back",
+ "A named storage location for a value."),
+ ("room_summaries.summary", "room_summaries", "room_id",
+ "rich-room-study-group", "summary",
+ "The group is reviewing recursion before the midterm."),
 ]
 
 # (label, id_value, column, expected_number)
@@ -123,6 +130,14 @@ def test_quiz_context_context_json_is_ciphertext_and_decrypts(db_conn):
  assert decoded["asked"] == 2
 
 
+def test_study_guides_content_is_ciphertext_and_decrypts(db_conn):
+ """#518: study_guides.content needs the JSON pair (encrypt_json/decrypt_json)."""
+ raw = _raw(db_conn, "study_guides", "id", "rich-guide-cs-f25-mid", "content")
+ assert isinstance(raw, str), "content stored as PLAINTEXT JSONB — encryption regressed"
+ decoded = decrypt_json(raw)
+ assert decoded["exam"] == "Midterm Exam"
+
+
 def test_documents_concept_notes_uses_json_encryption(db_conn):
  """documents.concept_notes is stored via encrypt_json (a list of concepts)."""
  raw = _raw(db_conn, "documents", "id", "rich-doc-cs-syllabus", "concept_notes")

backend/tests/test_flashcard_import_service.py

@@ -59,6 +59,28 @@ def test_filters_by_topic_when_offering_id_is_none(self):
  assert "offering_id" not in str(call_kwargs)
  assert "course_id" not in str(call_kwargs)
 
+ def test_dedup_matches_against_encrypted_existing_cards(self):
+ # #518: dedupe must be decrypt-aware — existing rows are ciphertext.
+ from services.encryption import encrypt
+ with patch("services.flashcard_import_service.table") as t:
+ t.return_value.select.return_value = [
+ {"front": encrypt("What is a variable?")}
+ ]
+ new = [{"front": "What is a variable?", "back": "A named storage location."}]
+ keep, skipped = svc.dedup_against_existing("u1", "c1", new)
+ assert keep == []
+ assert skipped == new
+
+ def test_dedup_still_matches_legacy_plaintext_rows(self):
+ # Pre-backfill rows are still plaintext — decrypt_if_present falls
+ # back to the raw value, so dedupe keeps working unchanged.
+ with patch("services.flashcard_import_service.table") as t:
+ t.return_value.select.return_value = [{"front": "What is a variable?"}]
+ new = [{"front": "What is a variable?", "back": "A named storage location."}]
+ keep, skipped = svc.dedup_against_existing("u1", "c1", new)
+ assert keep == []
+ assert skipped == new
+
 
 # ── check_rate_limit ─────────────────────────────────────────────────────────
 

CLAUDE.md

@@ -106,7 +106,7 @@ make explore # Chapter 2: bounded AI exploration of the ru
 
 ## Gotchas
 
-- Column-level encryption is on for sensitive columns: `user_profiles.name`/`first_name`/`last_name`/`bio`/`location` (these moved off `users` to `user_profiles` in the 0024 identity split), Google OAuth tokens, `messages.content`, `room_messages.text`, `sessions.summary_json`, `documents.summary` + `concept_notes` + `extracted_text` (the RAG OCR text added in 0030), `notes.title`/`body`/`last_summary`, `assignments.notes`/`points_possible`/`points_earned` (the enrollment-keyed gradebook table; points columns carry numeric semantics — use `decrypt_numeric` at read), `feedback.comment`/`topic` + `issue_reports.topic`/`description` (free-text user input, #520), and `quiz_attempts.questions_json`/`answers_json` + `quiz_context.context_json` (quiz performance data, #521; scalar analytics columns — score/total/difficulty/completed_at — stay plaintext). Helpers live in `backend/services/encryption.py`; use `encrypt_if_present` at write boundaries and `decrypt_if_present` / `decrypt_numeric` at read boundaries (including before injecting into AI prompts). `ENCRYPTION_KEY` must be set (32 bytes as 64 hex chars; generate via `python -c "import secrets; print(secrets.token_hex(32))"`). Deliberate exception: `newsletter_emails.email` stays plaintext (ADR 0026) — the UNIQUE constraint, lookup index, and both subscribe/allowlist upserts key on the value, and AES-GCM's per-call nonce breaks value equality.
+- Column-level encryption is on for sensitive columns: `user_profiles.name`/`first_name`/`last_name`/`bio`/`location` (these moved off `users` to `user_profiles` in the 0024 identity split), Google OAuth tokens, `messages.content`, `room_messages.text`, `sessions.summary_json`, `documents.summary` + `concept_notes` + `extracted_text` (the RAG OCR text added in 0030), `notes.title`/`body`/`last_summary`, `assignments.notes`/`points_possible`/`points_earned` (the enrollment-keyed gradebook table; points columns carry numeric semantics — use `decrypt_numeric` at read), `feedback.comment`/`topic` + `issue_reports.topic`/`description` (free-text user input, #520), `quiz_attempts.questions_json`/`answers_json` + `quiz_context.context_json` (quiz performance data, #521; scalar analytics columns — score/total/difficulty/completed_at — stay plaintext), and `flashcards.front`/`back` + `study_guides.content` + `room_summaries.summary` (derived content, #518; `study_guides.content` uses the JSON pair, `room_summaries.summary` keys its cache on the separate plaintext `member_hash` column so encryption doesn't affect cache-hit lookups). Helpers live in `backend/services/encryption.py`; use `encrypt_if_present` at write boundaries and `decrypt_if_present` / `decrypt_numeric` at read boundaries (including before injecting into AI prompts). `ENCRYPTION_KEY` must be set (32 bytes as 64 hex chars; generate via `python -c "import secrets; print(secrets.token_hex(32))"`). Deliberate exception: `newsletter_emails.email` stays plaintext (ADR 0026) — the UNIQUE constraint, lookup index, and both subscribe/allowlist upserts key on the value, and AES-GCM's per-call nonce breaks value equality.
 - Knowledge-graph mastery is now an append-only `node_mastery_events` table (replaced the `graph_nodes.mastery_events` JSON column in 0023); node/edge dedup is enforced by UNIQUE constraints. Don't read/write a `mastery_events` column.
 - Optional cross-worker cache (#97): `services/cache.py` wraps Redis and is **off by default** — with no `REDIS_URL` set it's a zero-overhead no-op and never fails a request (any Redis error → clean miss + warning). Currently backs the content-addressed OCR/extraction cache (`extraction_service.extract_text_from_file`, keyed on `sha256(file_bytes)` + engine). The `redis` dependency is only imported when `REDIS_URL` is set.
 - HTTP caching (#99): conditional GETs use `services/http_cache.py` (`make_etag`/`conditional`/`cached_json`). `Cache-Control` on these routes is **always `private`, never `public`** — the responses carry user-scoped, app-decrypted columns that must never be cached at a shared proxy/CDN. Derive the ETag from cheap change-keys (ids, `updated_at`, existing content hashes), not from the fully-built payload.

backend/db/backfill_encryption.py

@@ -302,6 +302,18 @@ def backfill_quiz_context(apply: bool) -> dict:
  return _encrypt_json_column("quiz_context", "context_json", pk="id", apply=apply)
 
 
+def backfill_flashcards(apply: bool) -> dict:
+ return _encrypt_text_column("flashcards", ["front", "back"], pk="id", apply=apply)
+
+
+def backfill_study_guides(apply: bool) -> dict:
+ return _encrypt_json_column("study_guides", "content", pk="id", apply=apply)
+
+
+def backfill_room_summaries(apply: bool) -> dict:
+ return _encrypt_text_column("room_summaries", ["summary"], pk="room_id", apply=apply)
+
+
 RUNNERS: dict[str, Callable[[bool], dict]] = {
  "users": backfill_users,
  "user_settings": backfill_user_settings,
@@ -315,6 +327,9 @@ def backfill_quiz_context(apply: bool) -> dict:
  "issue_reports": backfill_issue_reports,
  "quiz_attempts": backfill_quiz_attempts,
  "quiz_context": backfill_quiz_context,
+ "flashcards": backfill_flashcards,
+ "study_guides": backfill_study_guides,
+ "room_summaries": backfill_room_summaries,
 }
 
 

backend/db/seed_local_rich.py

@@ -553,7 +553,7 @@ def seed_notes_documents() -> None:
  )
 
 
-# (fc_id, user_id, offering_id, topic, front, back) — plaintext, grouped by topic.
+# (fc_id, user_id, offering_id, topic, front, back) — 🔒 front/back, grouped by topic.
 _FLASHCARDS = [
  ("rich-fc-cs-1", USER_ACTIVE, OFF_CS_F25, "CS Basics",
  "What is a variable?", "A named storage location for a value."),
@@ -579,8 +579,9 @@ def seed_flashcards() -> None:
  "user_id": user_id,
  "offering_id": off_id,
  "topic": topic,
- "front": front,
- "back": back,
+ # 🔒 front / back (#518)
+ "front": encrypt_if_present(front),
+ "back": encrypt_if_present(back),
  },
  )
 
@@ -625,11 +626,26 @@ def seed_study_guides() -> None:
  "offering_id": off_id,
  "exam_id": exam_id,
  "generated_at": generated_at,
- "content": content,
+ # 🔒 content (#518)
+ "content": encrypt_json(content),
  },
  )
 
 
+def seed_room_summaries() -> None:
+ # #518: room_summaries.summary is 🔒. PK is room_id (no id column), so this
+ # can't go through insert_if_absent.
+ if not table("room_summaries").select("room_id", filters={"room_id": f"eq.{ROOM_STUDY}"}):
+ table("room_summaries").insert({
+ "room_id": ROOM_STUDY,
+ "summary": encrypt_if_present("The group is reviewing recursion before the midterm."),
+ "member_hash": "rich-member-hash-v1",
+ })
+ h.record("room_summaries", created=True)
+ else:
+ h.record("room_summaries", created=False)
+
+
 # (qa_id, concept_node_id, difficulty, score, total, questions_json, answers_json, completed_at)
 _QUIZ_ATTEMPTS = [
  ("rich-qa-cs-variables-1", "rich-node-cs-variables", "easy", 9, 10,
@@ -766,6 +782,7 @@ def seed_sessions() -> None:
  "schools", "courses", "course_offerings", "users", "user_profiles", "user_roles",
  "enrollments", "graph_nodes", "graph_edges", "node_mastery_events",
  "gradebook_categories", "assignments", "rooms", "room_members", "room_messages",
+ "room_summaries",
  "notes", "documents", "flashcards", "study_guides", "quiz_attempts", "quiz_context",
  "sessions", "messages", "feedback", "issue_reports",
 ]
@@ -785,6 +802,7 @@ def main() -> None:
  seed_notes_documents()
  seed_flashcards()
  seed_study_guides()
+ seed_room_summaries()
  seed_quiz()
  seed_feedback()
  seed_sessions()

backend/e2e_oracles/gather.py

@@ -195,6 +195,10 @@ def run_counts(args: argparse.Namespace) -> tuple[list[Finding], int]:
  ("quiz_attempts", "id", "questions_json"),
  ("quiz_attempts", "id", "answers_json"),
  ("quiz_context", "id", "context_json"),
+ ("flashcards", "id", "front"),
+ ("flashcards", "id", "back"),
+ ("study_guides", "id", "content"),
+ ("room_summaries", "room_id", "summary"),
 )
 
 

backend/routes/flashcards.py

@@ -14,7 +14,7 @@
 from services.academics import resolve_offering, term_id_for_label
 from services.auth_guard import require_self, get_session_user_id
 from services.achievement_service import check_achievements
-from services.encryption import decrypt_if_present, decrypt_json
+from services.encryption import decrypt_if_present, decrypt_json, encrypt_if_present
 from services.flashcard_import_service import (
  dedup_against_existing,
  check_rate_limit,
@@ -236,8 +236,8 @@ def generate(body: GenerateFlashcardsBody, request: Request):
  "id": str(uuid.uuid4()),
  "user_id": body.user_id,
  "topic": body.topic,
- "front": c["front"],
- "back": c["back"],
+ "front": encrypt_if_present(c["front"]),
+ "back": encrypt_if_present(c["back"]),
  "times_reviewed": 0,
  "last_reviewed_at": None,
  "created_at": now,
@@ -261,8 +261,16 @@ def generate(body: GenerateFlashcardsBody, request: Request):
  except Exception:
  pass
 
+ # rows_to_insert holds the ciphertext just written; the response the
+ # frontend renders (freshly generated cards, before any list re-fetch)
+ # must carry plaintext front/back, not the encrypted insert payload.
+ response_cards = [
+ {**row, "front": decrypt_if_present(row["front"]), "back": decrypt_if_present(row["back"])}
+ for row in rows_to_insert
+ ]
+
  return {
- "flashcards": rows_to_insert,
+ "flashcards": response_cards,
  "context_used": {
  "documents_found": len(documents),
  "weak_concepts_found": len(weak_concepts),
@@ -291,6 +299,9 @@ def get_flashcards(
  "id,user_id,topic,offering_id,front,back,times_reviewed,last_rating,last_reviewed_at,created_at",
  filters=filters, order="created_at.desc"
  ) or []
+ for r in rows:
+ r["front"] = decrypt_if_present(r.get("front"))
+ r["back"] = decrypt_if_present(r.get("back"))
  if semester:
  # Term scoping (#141). This route is user-wide (no course id), so
  # the filter works on the cards' offering: cards from the selected
@@ -397,8 +408,8 @@ def import_commit(body: ImportCommitBody, request: Request):
  "user_id": body.user_id,
  "topic": body.topic,
  "offering_id": offering_id,
- "front": c["front"],
- "back": c["back"],
+ "front": encrypt_if_present(c["front"]),
+ "back": encrypt_if_present(c["back"]),
  "times_reviewed": 0,
  "last_reviewed_at": None,
  "created_at": now,

backend/routes/study_guide.py

@@ -4,6 +4,7 @@
 Study guide generation and caching.
 """
 
+import logging
 import uuid
 from datetime import datetime, timezone
 
@@ -24,10 +25,17 @@
 )
 from services.graph_service import get_courses as graph_get_courses
 from services.auth_guard import require_self
-from services.encryption import decrypt_if_present, decrypt_json
+from services.encryption import (
+ decrypt_if_present,
+ decrypt_json,
+ decrypt_json_column,
+ encrypt_json,
+)
 from services.http_cache import cached_json, conditional, make_etag
 from services.request_context import current_request_id
 
+logger = logging.getLogger(__name__)
+
 router = APIRouter()
 
 
@@ -198,7 +206,7 @@ def _generate_and_insert(user_id: str, offering_id: str, exam_id: str) -> dict:
  "offering_id": offering_id,
  "exam_id": exam_id,
  "generated_at": now,
- "content": content,
+ "content": encrypt_json(content),
  }
  table("study_guides").insert(row)
 
@@ -245,7 +253,14 @@ def get_cached_guides(user_id: str, request: Request):
 
  result = []
  for g in guides:
- content = g.get("content") or {}
+ try:
+ content = decrypt_json_column(g.get("content")) or {}
+ except Exception:
+ logger.warning(
+ "get_cached_guides: content decrypt failed for guide %s; degrading",
+ g.get("id"),
+ )
+ content = {}
  course_id = offering_to_course.get(g.get("offering_id"))
  result.append({
  "id": g["id"],
@@ -330,7 +345,7 @@ def get_guide(
  )
  if cached:
  row = cached[0]
- return {"guide": row["content"], "generated_at": row["generated_at"], "cached": True}
+ return {"guide": decrypt_json_column(row["content"]), "generated_at": row["generated_at"], "cached": True}
 
  result = _generate_and_insert(user_id, offering_id, exam_id)
  return {"guide": result["content"], "generated_at": result["generated_at"], "cached": False}

backend/services/flashcard_import_service.py

@@ -33,6 +33,7 @@
 from agents.flashcard import flashcard_agent
 from agents.usage import record_agent_usage
 from services import extraction_service
+from services.encryption import decrypt_if_present
 
 logger = logging.getLogger(__name__)
 
@@ -76,7 +77,9 @@ def dedup_against_existing(
  filters["topic"] = f"eq.{topic}"
 
  existing = table("flashcards").select("front", filters=filters) or []
- existing_norm = [_normalize(r.get("front", "")) for r in existing]
+ existing_norm = [
+ _normalize(decrypt_if_present(r.get("front", "")) or "") for r in existing
+ ]
 
  keep: list[Card] = []
  skipped: list[Card] = []

backend/services/social_cache_service.py

@@ -12,6 +12,7 @@
 from datetime import datetime, timezone
 
 from db.connection import table
+from services.encryption import decrypt_if_present, encrypt_if_present
 
 
 def _compute_hash(member_summaries: list[str]) -> str:
@@ -26,15 +27,15 @@ def get_cached_summary(room_id: str, member_summaries: list[str]) -> str | None:
  filters={"room_id": f"eq.{room_id}"},
  )
  if rows and rows[0]["member_hash"] == current_hash:
- return rows[0]["summary"]
+ return decrypt_if_present(rows[0]["summary"])
  return None
 
 
 def save_summary(room_id: str, member_summaries: list[str], summary: str) -> None:
  table("room_summaries").upsert(
  {
  "room_id": room_id,
- "summary": summary,
+ "summary": encrypt_if_present(summary),
  "member_hash": _compute_hash(member_summaries),
  "updated_at": datetime.now(timezone.utc).isoformat(),
  },

backend/tests/integration/test_encryption_roundtrip.py

@@ -46,6 +46,13 @@
  "Upload stuck"),
  ("issue_reports.description", "issue_reports", "id", "rich-issue-1",
  "description", "Syllabus upload spins forever."),
+ ("flashcards.front", "flashcards", "id", "rich-fc-cs-1", "front",
+ "What is a variable?"),
+ ("flashcards.back", "flashcards", "id", "rich-fc-cs-1", "back",
+ "A named storage location for a value."),
+ ("room_summaries.summary", "room_summaries", "room_id",
+ "rich-room-study-group", "summary",
+ "The group is reviewing recursion before the midterm."),
 ]
 
 # (label, id_value, column, expected_number)
@@ -123,6 +130,14 @@ def test_quiz_context_context_json_is_ciphertext_and_decrypts(db_conn):
  assert decoded["asked"] == 2
 
 
+def test_study_guides_content_is_ciphertext_and_decrypts(db_conn):
+ """#518: study_guides.content needs the JSON pair (encrypt_json/decrypt_json)."""
+ raw = _raw(db_conn, "study_guides", "id", "rich-guide-cs-f25-mid", "content")
+ assert isinstance(raw, str), "content stored as PLAINTEXT JSONB — encryption regressed"
+ decoded = decrypt_json(raw)
+ assert decoded["exam"] == "Midterm Exam"
+
+
 def test_documents_concept_notes_uses_json_encryption(db_conn):
  """documents.concept_notes is stored via encrypt_json (a list of concepts)."""
  raw = _raw(db_conn, "documents", "id", "rich-doc-cs-syllabus", "concept_notes")

backend/tests/test_flashcard_import_service.py

@@ -59,6 +59,28 @@ def test_filters_by_topic_when_offering_id_is_none(self):
  assert "offering_id" not in str(call_kwargs)
  assert "course_id" not in str(call_kwargs)
 
+ def test_dedup_matches_against_encrypted_existing_cards(self):
+ # #518: dedupe must be decrypt-aware — existing rows are ciphertext.
+ from services.encryption import encrypt
+ with patch("services.flashcard_import_service.table") as t:
+ t.return_value.select.return_value = [
+ {"front": encrypt("What is a variable?")}
+ ]
+ new = [{"front": "What is a variable?", "back": "A named storage location."}]
+ keep, skipped = svc.dedup_against_existing("u1", "c1", new)
+ assert keep == []
+ assert skipped == new
+
+ def test_dedup_still_matches_legacy_plaintext_rows(self):
+ # Pre-backfill rows are still plaintext — decrypt_if_present falls
+ # back to the raw value, so dedupe keeps working unchanged.
+ with patch("services.flashcard_import_service.table") as t:
+ t.return_value.select.return_value = [{"front": "What is a variable?"}]
+ new = [{"front": "What is a variable?", "back": "A named storage location."}]
+ keep, skipped = svc.dedup_against_existing("u1", "c1", new)
+ assert keep == []
+ assert skipped == new
+
 
 # ── check_rate_limit ─────────────────────────────────────────────────────────
 

CLAUDE.md

@@ -106,7 +106,7 @@ make explore # Chapter 2: bounded AI exploration of the ru
 
 ## Gotchas
 
-- Column-level encryption is on for sensitive columns: `user_profiles.name`/`first_name`/`last_name`/`bio`/`location` (these moved off `users` to `user_profiles` in the 0024 identity split), Google OAuth tokens, `messages.content`, `room_messages.text`, `sessions.summary_json`, `documents.summary` + `concept_notes` + `extracted_text` (the RAG OCR text added in 0030), `notes.title`/`body`/`last_summary`, `assignments.notes`/`points_possible`/`points_earned` (the enrollment-keyed gradebook table; points columns carry numeric semantics — use `decrypt_numeric` at read), `feedback.comment`/`topic` + `issue_reports.topic`/`description` (free-text user input, #520), and `quiz_attempts.questions_json`/`answers_json` + `quiz_context.context_json` (quiz performance data, #521; scalar analytics columns — score/total/difficulty/completed_at — stay plaintext). Helpers live in `backend/services/encryption.py`; use `encrypt_if_present` at write boundaries and `decrypt_if_present` / `decrypt_numeric` at read boundaries (including before injecting into AI prompts). `ENCRYPTION_KEY` must be set (32 bytes as 64 hex chars; generate via `python -c "import secrets; print(secrets.token_hex(32))"`). Deliberate exception: `newsletter_emails.email` stays plaintext (ADR 0026) — the UNIQUE constraint, lookup index, and both subscribe/allowlist upserts key on the value, and AES-GCM's per-call nonce breaks value equality.
+- Column-level encryption is on for sensitive columns: `user_profiles.name`/`first_name`/`last_name`/`bio`/`location` (these moved off `users` to `user_profiles` in the 0024 identity split), Google OAuth tokens, `messages.content`, `room_messages.text`, `sessions.summary_json`, `documents.summary` + `concept_notes` + `extracted_text` (the RAG OCR text added in 0030), `notes.title`/`body`/`last_summary`, `assignments.notes`/`points_possible`/`points_earned` (the enrollment-keyed gradebook table; points columns carry numeric semantics — use `decrypt_numeric` at read), `feedback.comment`/`topic` + `issue_reports.topic`/`description` (free-text user input, #520), `quiz_attempts.questions_json`/`answers_json` + `quiz_context.context_json` (quiz performance data, #521; scalar analytics columns — score/total/difficulty/completed_at — stay plaintext), and `flashcards.front`/`back` + `study_guides.content` + `room_summaries.summary` (derived content, #518; `study_guides.content` uses the JSON pair, `room_summaries.summary` keys its cache on the separate plaintext `member_hash` column so encryption doesn't affect cache-hit lookups). Helpers live in `backend/services/encryption.py`; use `encrypt_if_present` at write boundaries and `decrypt_if_present` / `decrypt_numeric` at read boundaries (including before injecting into AI prompts). `ENCRYPTION_KEY` must be set (32 bytes as 64 hex chars; generate via `python -c "import secrets; print(secrets.token_hex(32))"`). Deliberate exception: `newsletter_emails.email` stays plaintext (ADR 0026) — the UNIQUE constraint, lookup index, and both subscribe/allowlist upserts key on the value, and AES-GCM's per-call nonce breaks value equality.
 - Knowledge-graph mastery is now an append-only `node_mastery_events` table (replaced the `graph_nodes.mastery_events` JSON column in 0023); node/edge dedup is enforced by UNIQUE constraints. Don't read/write a `mastery_events` column.
 - Optional cross-worker cache (#97): `services/cache.py` wraps Redis and is **off by default** — with no `REDIS_URL` set it's a zero-overhead no-op and never fails a request (any Redis error → clean miss + warning). Currently backs the content-addressed OCR/extraction cache (`extraction_service.extract_text_from_file`, keyed on `sha256(file_bytes)` + engine). The `redis` dependency is only imported when `REDIS_URL` is set.
 - HTTP caching (#99): conditional GETs use `services/http_cache.py` (`make_etag`/`conditional`/`cached_json`). `Cache-Control` on these routes is **always `private`, never `public`** — the responses carry user-scoped, app-decrypted columns that must never be cached at a shared proxy/CDN. Derive the ETag from cheap change-keys (ids, `updated_at`, existing content hashes), not from the fully-built payload.

backend/db/backfill_encryption.py

@@ -302,6 +302,18 @@ def backfill_quiz_context(apply: bool) -> dict:
  return _encrypt_json_column("quiz_context", "context_json", pk="id", apply=apply)
 
 
+def backfill_flashcards(apply: bool) -> dict:
+ return _encrypt_text_column("flashcards", ["front", "back"], pk="id", apply=apply)
+
+
+def backfill_study_guides(apply: bool) -> dict:
+ return _encrypt_json_column("study_guides", "content", pk="id", apply=apply)
+
+
+def backfill_room_summaries(apply: bool) -> dict:
+ return _encrypt_text_column("room_summaries", ["summary"], pk="room_id", apply=apply)
+
+
 RUNNERS: dict[str, Callable[[bool], dict]] = {
  "users": backfill_users,
  "user_settings": backfill_user_settings,
@@ -315,6 +327,9 @@ def backfill_quiz_context(apply: bool) -> dict:
  "issue_reports": backfill_issue_reports,
  "quiz_attempts": backfill_quiz_attempts,
  "quiz_context": backfill_quiz_context,
+ "flashcards": backfill_flashcards,
+ "study_guides": backfill_study_guides,
+ "room_summaries": backfill_room_summaries,
 }
 
 

backend/db/seed_local_rich.py

@@ -553,7 +553,7 @@ def seed_notes_documents() -> None:
  )
 
 
-# (fc_id, user_id, offering_id, topic, front, back) — plaintext, grouped by topic.
+# (fc_id, user_id, offering_id, topic, front, back) — 🔒 front/back, grouped by topic.
 _FLASHCARDS = [
  ("rich-fc-cs-1", USER_ACTIVE, OFF_CS_F25, "CS Basics",
  "What is a variable?", "A named storage location for a value."),
@@ -579,8 +579,9 @@ def seed_flashcards() -> None:
  "user_id": user_id,
  "offering_id": off_id,
  "topic": topic,
- "front": front,
- "back": back,
+ # 🔒 front / back (#518)
+ "front": encrypt_if_present(front),
+ "back": encrypt_if_present(back),
  },
  )
 
@@ -625,11 +626,26 @@ def seed_study_guides() -> None:
  "offering_id": off_id,
  "exam_id": exam_id,
  "generated_at": generated_at,
- "content": content,
+ # 🔒 content (#518)
+ "content": encrypt_json(content),
  },
  )
 
 
+def seed_room_summaries() -> None:
+ # #518: room_summaries.summary is 🔒. PK is room_id (no id column), so this
+ # can't go through insert_if_absent.
+ if not table("room_summaries").select("room_id", filters={"room_id": f"eq.{ROOM_STUDY}"}):
+ table("room_summaries").insert({
+ "room_id": ROOM_STUDY,
+ "summary": encrypt_if_present("The group is reviewing recursion before the midterm."),
+ "member_hash": "rich-member-hash-v1",
+ })
+ h.record("room_summaries", created=True)
+ else:
+ h.record("room_summaries", created=False)
+
+
 # (qa_id, concept_node_id, difficulty, score, total, questions_json, answers_json, completed_at)
 _QUIZ_ATTEMPTS = [
  ("rich-qa-cs-variables-1", "rich-node-cs-variables", "easy", 9, 10,
@@ -766,6 +782,7 @@ def seed_sessions() -> None:
  "schools", "courses", "course_offerings", "users", "user_profiles", "user_roles",
  "enrollments", "graph_nodes", "graph_edges", "node_mastery_events",
  "gradebook_categories", "assignments", "rooms", "room_members", "room_messages",
+ "room_summaries",
  "notes", "documents", "flashcards", "study_guides", "quiz_attempts", "quiz_context",
  "sessions", "messages", "feedback", "issue_reports",
 ]
@@ -785,6 +802,7 @@ def main() -> None:
  seed_notes_documents()
  seed_flashcards()
  seed_study_guides()
+ seed_room_summaries()
  seed_quiz()
  seed_feedback()
  seed_sessions()

backend/e2e_oracles/gather.py

@@ -195,6 +195,10 @@ def run_counts(args: argparse.Namespace) -> tuple[list[Finding], int]:
  ("quiz_attempts", "id", "questions_json"),
  ("quiz_attempts", "id", "answers_json"),
  ("quiz_context", "id", "context_json"),
+ ("flashcards", "id", "front"),
+ ("flashcards", "id", "back"),
+ ("study_guides", "id", "content"),
+ ("room_summaries", "room_id", "summary"),
 )
 
 

backend/routes/flashcards.py

@@ -14,7 +14,7 @@
 from services.academics import resolve_offering, term_id_for_label
 from services.auth_guard import require_self, get_session_user_id
 from services.achievement_service import check_achievements
-from services.encryption import decrypt_if_present, decrypt_json
+from services.encryption import decrypt_if_present, decrypt_json, encrypt_if_present
 from services.flashcard_import_service import (
  dedup_against_existing,
  check_rate_limit,
@@ -236,8 +236,8 @@ def generate(body: GenerateFlashcardsBody, request: Request):
  "id": str(uuid.uuid4()),
  "user_id": body.user_id,
  "topic": body.topic,
- "front": c["front"],
- "back": c["back"],
+ "front": encrypt_if_present(c["front"]),
+ "back": encrypt_if_present(c["back"]),
  "times_reviewed": 0,
  "last_reviewed_at": None,
  "created_at": now,
@@ -261,8 +261,16 @@ def generate(body: GenerateFlashcardsBody, request: Request):
  except Exception:
  pass
 
+ # rows_to_insert holds the ciphertext just written; the response the
+ # frontend renders (freshly generated cards, before any list re-fetch)
+ # must carry plaintext front/back, not the encrypted insert payload.
+ response_cards = [
+ {**row, "front": decrypt_if_present(row["front"]), "back": decrypt_if_present(row["back"])}
+ for row in rows_to_insert
+ ]
+
  return {
- "flashcards": rows_to_insert,
+ "flashcards": response_cards,
  "context_used": {
  "documents_found": len(documents),
  "weak_concepts_found": len(weak_concepts),
@@ -291,6 +299,9 @@ def get_flashcards(
  "id,user_id,topic,offering_id,front,back,times_reviewed,last_rating,last_reviewed_at,created_at",
  filters=filters, order="created_at.desc"
  ) or []
+ for r in rows:
+ r["front"] = decrypt_if_present(r.get("front"))
+ r["back"] = decrypt_if_present(r.get("back"))
  if semester:
  # Term scoping (#141). This route is user-wide (no course id), so
  # the filter works on the cards' offering: cards from the selected
@@ -397,8 +408,8 @@ def import_commit(body: ImportCommitBody, request: Request):
  "user_id": body.user_id,
  "topic": body.topic,
  "offering_id": offering_id,
- "front": c["front"],
- "back": c["back"],
+ "front": encrypt_if_present(c["front"]),
+ "back": encrypt_if_present(c["back"]),
  "times_reviewed": 0,
  "last_reviewed_at": None,
  "created_at": now,

backend/routes/study_guide.py

@@ -4,6 +4,7 @@
 Study guide generation and caching.
 """
 
+import logging
 import uuid
 from datetime import datetime, timezone
 
@@ -24,10 +25,17 @@
 )
 from services.graph_service import get_courses as graph_get_courses
 from services.auth_guard import require_self
-from services.encryption import decrypt_if_present, decrypt_json
+from services.encryption import (
+ decrypt_if_present,
+ decrypt_json,
+ decrypt_json_column,
+ encrypt_json,
+)
 from services.http_cache import cached_json, conditional, make_etag
 from services.request_context import current_request_id
 
+logger = logging.getLogger(__name__)
+
 router = APIRouter()
 
 
@@ -198,7 +206,7 @@ def _generate_and_insert(user_id: str, offering_id: str, exam_id: str) -> dict:
  "offering_id": offering_id,
  "exam_id": exam_id,
  "generated_at": now,
- "content": content,
+ "content": encrypt_json(content),
  }
  table("study_guides").insert(row)
 
@@ -245,7 +253,14 @@ def get_cached_guides(user_id: str, request: Request):
 
  result = []
  for g in guides:
- content = g.get("content") or {}
+ try:
+ content = decrypt_json_column(g.get("content")) or {}
+ except Exception:
+ logger.warning(
+ "get_cached_guides: content decrypt failed for guide %s; degrading",
+ g.get("id"),
+ )
+ content = {}
  course_id = offering_to_course.get(g.get("offering_id"))
  result.append({
  "id": g["id"],
@@ -330,7 +345,7 @@ def get_guide(
  )
  if cached:
  row = cached[0]
- return {"guide": row["content"], "generated_at": row["generated_at"], "cached": True}
+ return {"guide": decrypt_json_column(row["content"]), "generated_at": row["generated_at"], "cached": True}
 
  result = _generate_and_insert(user_id, offering_id, exam_id)
  return {"guide": result["content"], "generated_at": result["generated_at"], "cached": False}

backend/services/flashcard_import_service.py

@@ -33,6 +33,7 @@
 from agents.flashcard import flashcard_agent
 from agents.usage import record_agent_usage
 from services import extraction_service
+from services.encryption import decrypt_if_present
 
 logger = logging.getLogger(__name__)
 
@@ -76,7 +77,9 @@ def dedup_against_existing(
  filters["topic"] = f"eq.{topic}"
 
  existing = table("flashcards").select("front", filters=filters) or []
- existing_norm = [_normalize(r.get("front", "")) for r in existing]
+ existing_norm = [
+ _normalize(decrypt_if_present(r.get("front", "")) or "") for r in existing
+ ]
 
  keep: list[Card] = []
  skipped: list[Card] = []

backend/services/social_cache_service.py

@@ -12,6 +12,7 @@
 from datetime import datetime, timezone
 
 from db.connection import table
+from services.encryption import decrypt_if_present, encrypt_if_present
 
 
 def _compute_hash(member_summaries: list[str]) -> str:
@@ -26,15 +27,15 @@ def get_cached_summary(room_id: str, member_summaries: list[str]) -> str | None:
  filters={"room_id": f"eq.{room_id}"},
  )
  if rows and rows[0]["member_hash"] == current_hash:
- return rows[0]["summary"]
+ return decrypt_if_present(rows[0]["summary"])
  return None
 
 
 def save_summary(room_id: str, member_summaries: list[str], summary: str) -> None:
  table("room_summaries").upsert(
  {
  "room_id": room_id,
- "summary": summary,
+ "summary": encrypt_if_present(summary),
  "member_hash": _compute_hash(member_summaries),
  "updated_at": datetime.now(timezone.utc).isoformat(),
  },

backend/tests/integration/test_encryption_roundtrip.py

@@ -46,6 +46,13 @@
  "Upload stuck"),
  ("issue_reports.description", "issue_reports", "id", "rich-issue-1",
  "description", "Syllabus upload spins forever."),
+ ("flashcards.front", "flashcards", "id", "rich-fc-cs-1", "front",
+ "What is a variable?"),
+ ("flashcards.back", "flashcards", "id", "rich-fc-cs-1", "back",
+ "A named storage location for a value."),
+ ("room_summaries.summary", "room_summaries", "room_id",
+ "rich-room-study-group", "summary",
+ "The group is reviewing recursion before the midterm."),
 ]
 
 # (label, id_value, column, expected_number)
@@ -123,6 +130,14 @@ def test_quiz_context_context_json_is_ciphertext_and_decrypts(db_conn):
  assert decoded["asked"] == 2
 
 
+def test_study_guides_content_is_ciphertext_and_decrypts(db_conn):
+ """#518: study_guides.content needs the JSON pair (encrypt_json/decrypt_json)."""
+ raw = _raw(db_conn, "study_guides", "id", "rich-guide-cs-f25-mid", "content")
+ assert isinstance(raw, str), "content stored as PLAINTEXT JSONB — encryption regressed"
+ decoded = decrypt_json(raw)
+ assert decoded["exam"] == "Midterm Exam"
+
+
 def test_documents_concept_notes_uses_json_encryption(db_conn):
  """documents.concept_notes is stored via encrypt_json (a list of concepts)."""
  raw = _raw(db_conn, "documents", "id", "rich-doc-cs-syllabus", "concept_notes")

backend/tests/test_flashcard_import_service.py

@@ -59,6 +59,28 @@ def test_filters_by_topic_when_offering_id_is_none(self):
  assert "offering_id" not in str(call_kwargs)
  assert "course_id" not in str(call_kwargs)
 
+ def test_dedup_matches_against_encrypted_existing_cards(self):
+ # #518: dedupe must be decrypt-aware — existing rows are ciphertext.
+ from services.encryption import encrypt
+ with patch("services.flashcard_import_service.table") as t:
+ t.return_value.select.return_value = [
+ {"front": encrypt("What is a variable?")}
+ ]
+ new = [{"front": "What is a variable?", "back": "A named storage location."}]
+ keep, skipped = svc.dedup_against_existing("u1", "c1", new)
+ assert keep == []
+ assert skipped == new
+
+ def test_dedup_still_matches_legacy_plaintext_rows(self):
+ # Pre-backfill rows are still plaintext — decrypt_if_present falls
+ # back to the raw value, so dedupe keeps working unchanged.
+ with patch("services.flashcard_import_service.table") as t:
+ t.return_value.select.return_value = [{"front": "What is a variable?"}]
+ new = [{"front": "What is a variable?", "back": "A named storage location."}]
+ keep, skipped = svc.dedup_against_existing("u1", "c1", new)
+ assert keep == []
+ assert skipped == new
+
 
 # ── check_rate_limit ─────────────────────────────────────────────────────────
 

CLAUDE.md

@@ -106,7 +106,7 @@ make explore # Chapter 2: bounded AI exploration of the ru
 
 ## Gotchas
 
-- Column-level encryption is on for sensitive columns: `user_profiles.name`/`first_name`/`last_name`/`bio`/`location` (these moved off `users` to `user_profiles` in the 0024 identity split), Google OAuth tokens, `messages.content`, `room_messages.text`, `sessions.summary_json`, `documents.summary` + `concept_notes` + `extracted_text` (the RAG OCR text added in 0030), `notes.title`/`body`/`last_summary`, `assignments.notes`/`points_possible`/`points_earned` (the enrollment-keyed gradebook table; points columns carry numeric semantics — use `decrypt_numeric` at read), `feedback.comment`/`topic` + `issue_reports.topic`/`description` (free-text user input, #520), and `quiz_attempts.questions_json`/`answers_json` + `quiz_context.context_json` (quiz performance data, #521; scalar analytics columns — score/total/difficulty/completed_at — stay plaintext). Helpers live in `backend/services/encryption.py`; use `encrypt_if_present` at write boundaries and `decrypt_if_present` / `decrypt_numeric` at read boundaries (including before injecting into AI prompts). `ENCRYPTION_KEY` must be set (32 bytes as 64 hex chars; generate via `python -c "import secrets; print(secrets.token_hex(32))"`). Deliberate exception: `newsletter_emails.email` stays plaintext (ADR 0026) — the UNIQUE constraint, lookup index, and both subscribe/allowlist upserts key on the value, and AES-GCM's per-call nonce breaks value equality.
+- Column-level encryption is on for sensitive columns: `user_profiles.name`/`first_name`/`last_name`/`bio`/`location` (these moved off `users` to `user_profiles` in the 0024 identity split), Google OAuth tokens, `messages.content`, `room_messages.text`, `sessions.summary_json`, `documents.summary` + `concept_notes` + `extracted_text` (the RAG OCR text added in 0030), `notes.title`/`body`/`last_summary`, `assignments.notes`/`points_possible`/`points_earned` (the enrollment-keyed gradebook table; points columns carry numeric semantics — use `decrypt_numeric` at read), `feedback.comment`/`topic` + `issue_reports.topic`/`description` (free-text user input, #520), `quiz_attempts.questions_json`/`answers_json` + `quiz_context.context_json` (quiz performance data, #521; scalar analytics columns — score/total/difficulty/completed_at — stay plaintext), and `flashcards.front`/`back` + `study_guides.content` + `room_summaries.summary` (derived content, #518; `study_guides.content` uses the JSON pair, `room_summaries.summary` keys its cache on the separate plaintext `member_hash` column so encryption doesn't affect cache-hit lookups). Helpers live in `backend/services/encryption.py`; use `encrypt_if_present` at write boundaries and `decrypt_if_present` / `decrypt_numeric` at read boundaries (including before injecting into AI prompts). `ENCRYPTION_KEY` must be set (32 bytes as 64 hex chars; generate via `python -c "import secrets; print(secrets.token_hex(32))"`). Deliberate exception: `newsletter_emails.email` stays plaintext (ADR 0026) — the UNIQUE constraint, lookup index, and both subscribe/allowlist upserts key on the value, and AES-GCM's per-call nonce breaks value equality.
 - Knowledge-graph mastery is now an append-only `node_mastery_events` table (replaced the `graph_nodes.mastery_events` JSON column in 0023); node/edge dedup is enforced by UNIQUE constraints. Don't read/write a `mastery_events` column.
 - Optional cross-worker cache (#97): `services/cache.py` wraps Redis and is **off by default** — with no `REDIS_URL` set it's a zero-overhead no-op and never fails a request (any Redis error → clean miss + warning). Currently backs the content-addressed OCR/extraction cache (`extraction_service.extract_text_from_file`, keyed on `sha256(file_bytes)` + engine). The `redis` dependency is only imported when `REDIS_URL` is set.
 - HTTP caching (#99): conditional GETs use `services/http_cache.py` (`make_etag`/`conditional`/`cached_json`). `Cache-Control` on these routes is **always `private`, never `public`** — the responses carry user-scoped, app-decrypted columns that must never be cached at a shared proxy/CDN. Derive the ETag from cheap change-keys (ids, `updated_at`, existing content hashes), not from the fully-built payload.

backend/db/backfill_encryption.py

@@ -302,6 +302,18 @@ def backfill_quiz_context(apply: bool) -> dict:
  return _encrypt_json_column("quiz_context", "context_json", pk="id", apply=apply)
 
 
+def backfill_flashcards(apply: bool) -> dict:
+ return _encrypt_text_column("flashcards", ["front", "back"], pk="id", apply=apply)
+
+
+def backfill_study_guides(apply: bool) -> dict:
+ return _encrypt_json_column("study_guides", "content", pk="id", apply=apply)
+
+
+def backfill_room_summaries(apply: bool) -> dict:
+ return _encrypt_text_column("room_summaries", ["summary"], pk="room_id", apply=apply)
+
+
 RUNNERS: dict[str, Callable[[bool], dict]] = {
  "users": backfill_users,
  "user_settings": backfill_user_settings,
@@ -315,6 +327,9 @@ def backfill_quiz_context(apply: bool) -> dict:
  "issue_reports": backfill_issue_reports,
  "quiz_attempts": backfill_quiz_attempts,
  "quiz_context": backfill_quiz_context,
+ "flashcards": backfill_flashcards,
+ "study_guides": backfill_study_guides,
+ "room_summaries": backfill_room_summaries,
 }
 
 

backend/db/seed_local_rich.py

@@ -553,7 +553,7 @@ def seed_notes_documents() -> None:
  )
 
 
-# (fc_id, user_id, offering_id, topic, front, back) — plaintext, grouped by topic.
+# (fc_id, user_id, offering_id, topic, front, back) — 🔒 front/back, grouped by topic.
 _FLASHCARDS = [
  ("rich-fc-cs-1", USER_ACTIVE, OFF_CS_F25, "CS Basics",
  "What is a variable?", "A named storage location for a value."),
@@ -579,8 +579,9 @@ def seed_flashcards() -> None:
  "user_id": user_id,
  "offering_id": off_id,
  "topic": topic,
- "front": front,
- "back": back,
+ # 🔒 front / back (#518)
+ "front": encrypt_if_present(front),
+ "back": encrypt_if_present(back),
  },
  )
 
@@ -625,11 +626,26 @@ def seed_study_guides() -> None:
  "offering_id": off_id,
  "exam_id": exam_id,
  "generated_at": generated_at,
- "content": content,
+ # 🔒 content (#518)
+ "content": encrypt_json(content),
  },
  )
 
 
+def seed_room_summaries() -> None:
+ # #518: room_summaries.summary is 🔒. PK is room_id (no id column), so this
+ # can't go through insert_if_absent.
+ if not table("room_summaries").select("room_id", filters={"room_id": f"eq.{ROOM_STUDY}"}):
+ table("room_summaries").insert({
+ "room_id": ROOM_STUDY,
+ "summary": encrypt_if_present("The group is reviewing recursion before the midterm."),
+ "member_hash": "rich-member-hash-v1",
+ })
+ h.record("room_summaries", created=True)
+ else:
+ h.record("room_summaries", created=False)
+
+
 # (qa_id, concept_node_id, difficulty, score, total, questions_json, answers_json, completed_at)
 _QUIZ_ATTEMPTS = [
  ("rich-qa-cs-variables-1", "rich-node-cs-variables", "easy", 9, 10,
@@ -766,6 +782,7 @@ def seed_sessions() -> None:
  "schools", "courses", "course_offerings", "users", "user_profiles", "user_roles",
  "enrollments", "graph_nodes", "graph_edges", "node_mastery_events",
  "gradebook_categories", "assignments", "rooms", "room_members", "room_messages",
+ "room_summaries",
  "notes", "documents", "flashcards", "study_guides", "quiz_attempts", "quiz_context",
  "sessions", "messages", "feedback", "issue_reports",
 ]
@@ -785,6 +802,7 @@ def main() -> None:
  seed_notes_documents()
  seed_flashcards()
  seed_study_guides()
+ seed_room_summaries()
  seed_quiz()
  seed_feedback()
  seed_sessions()

backend/e2e_oracles/gather.py

@@ -195,6 +195,10 @@ def run_counts(args: argparse.Namespace) -> tuple[list[Finding], int]:
  ("quiz_attempts", "id", "questions_json"),
  ("quiz_attempts", "id", "answers_json"),
  ("quiz_context", "id", "context_json"),
+ ("flashcards", "id", "front"),
+ ("flashcards", "id", "back"),
+ ("study_guides", "id", "content"),
+ ("room_summaries", "room_id", "summary"),
 )
 
 

backend/routes/flashcards.py

@@ -14,7 +14,7 @@
 from services.academics import resolve_offering, term_id_for_label
 from services.auth_guard import require_self, get_session_user_id
 from services.achievement_service import check_achievements
-from services.encryption import decrypt_if_present, decrypt_json
+from services.encryption import decrypt_if_present, decrypt_json, encrypt_if_present
 from services.flashcard_import_service import (
  dedup_against_existing,
  check_rate_limit,
@@ -236,8 +236,8 @@ def generate(body: GenerateFlashcardsBody, request: Request):
  "id": str(uuid.uuid4()),
  "user_id": body.user_id,
  "topic": body.topic,
- "front": c["front"],
- "back": c["back"],
+ "front": encrypt_if_present(c["front"]),
+ "back": encrypt_if_present(c["back"]),
  "times_reviewed": 0,
  "last_reviewed_at": None,
  "created_at": now,
@@ -261,8 +261,16 @@ def generate(body: GenerateFlashcardsBody, request: Request):
  except Exception:
  pass
 
+ # rows_to_insert holds the ciphertext just written; the response the
+ # frontend renders (freshly generated cards, before any list re-fetch)
+ # must carry plaintext front/back, not the encrypted insert payload.
+ response_cards = [
+ {**row, "front": decrypt_if_present(row["front"]), "back": decrypt_if_present(row["back"])}
+ for row in rows_to_insert
+ ]
+
  return {
- "flashcards": rows_to_insert,
+ "flashcards": response_cards,
  "context_used": {
  "documents_found": len(documents),
  "weak_concepts_found": len(weak_concepts),
@@ -291,6 +299,9 @@ def get_flashcards(
  "id,user_id,topic,offering_id,front,back,times_reviewed,last_rating,last_reviewed_at,created_at",
  filters=filters, order="created_at.desc"
  ) or []
+ for r in rows:
+ r["front"] = decrypt_if_present(r.get("front"))
+ r["back"] = decrypt_if_present(r.get("back"))
  if semester:
  # Term scoping (#141). This route is user-wide (no course id), so
  # the filter works on the cards' offering: cards from the selected
@@ -397,8 +408,8 @@ def import_commit(body: ImportCommitBody, request: Request):
  "user_id": body.user_id,
  "topic": body.topic,
  "offering_id": offering_id,
- "front": c["front"],
- "back": c["back"],
+ "front": encrypt_if_present(c["front"]),
+ "back": encrypt_if_present(c["back"]),
  "times_reviewed": 0,
  "last_reviewed_at": None,
  "created_at": now,

backend/routes/study_guide.py

@@ -4,6 +4,7 @@
 Study guide generation and caching.
 """
 
+import logging
 import uuid
 from datetime import datetime, timezone
 
@@ -24,10 +25,17 @@
 )
 from services.graph_service import get_courses as graph_get_courses
 from services.auth_guard import require_self
-from services.encryption import decrypt_if_present, decrypt_json
+from services.encryption import (
+ decrypt_if_present,
+ decrypt_json,
+ decrypt_json_column,
+ encrypt_json,
+)
 from services.http_cache import cached_json, conditional, make_etag
 from services.request_context import current_request_id
 
+logger = logging.getLogger(__name__)
+
 router = APIRouter()
 
 
@@ -198,7 +206,7 @@ def _generate_and_insert(user_id: str, offering_id: str, exam_id: str) -> dict:
  "offering_id": offering_id,
  "exam_id": exam_id,
  "generated_at": now,
- "content": content,
+ "content": encrypt_json(content),
  }
  table("study_guides").insert(row)
 
@@ -245,7 +253,14 @@ def get_cached_guides(user_id: str, request: Request):
 
  result = []
  for g in guides:
- content = g.get("content") or {}
+ try:
+ content = decrypt_json_column(g.get("content")) or {}
+ except Exception:
+ logger.warning(
+ "get_cached_guides: content decrypt failed for guide %s; degrading",
+ g.get("id"),
+ )
+ content = {}
  course_id = offering_to_course.get(g.get("offering_id"))
  result.append({
  "id": g["id"],
@@ -330,7 +345,7 @@ def get_guide(
  )
  if cached:
  row = cached[0]
- return {"guide": row["content"], "generated_at": row["generated_at"], "cached": True}
+ return {"guide": decrypt_json_column(row["content"]), "generated_at": row["generated_at"], "cached": True}
 
  result = _generate_and_insert(user_id, offering_id, exam_id)
  return {"guide": result["content"], "generated_at": result["generated_at"], "cached": False}

backend/services/flashcard_import_service.py

@@ -33,6 +33,7 @@
 from agents.flashcard import flashcard_agent
 from agents.usage import record_agent_usage
 from services import extraction_service
+from services.encryption import decrypt_if_present
 
 logger = logging.getLogger(__name__)
 
@@ -76,7 +77,9 @@ def dedup_against_existing(
  filters["topic"] = f"eq.{topic}"
 
  existing = table("flashcards").select("front", filters=filters) or []
- existing_norm = [_normalize(r.get("front", "")) for r in existing]
+ existing_norm = [
+ _normalize(decrypt_if_present(r.get("front", "")) or "") for r in existing
+ ]
 
  keep: list[Card] = []
  skipped: list[Card] = []

backend/services/social_cache_service.py

@@ -12,6 +12,7 @@
 from datetime import datetime, timezone
 
 from db.connection import table
+from services.encryption import decrypt_if_present, encrypt_if_present
 
 
 def _compute_hash(member_summaries: list[str]) -> str:
@@ -26,15 +27,15 @@ def get_cached_summary(room_id: str, member_summaries: list[str]) -> str | None:
  filters={"room_id": f"eq.{room_id}"},
  )
  if rows and rows[0]["member_hash"] == current_hash:
- return rows[0]["summary"]
+ return decrypt_if_present(rows[0]["summary"])
  return None
 
 
 def save_summary(room_id: str, member_summaries: list[str], summary: str) -> None:
  table("room_summaries").upsert(
  {
  "room_id": room_id,
- "summary": summary,
+ "summary": encrypt_if_present(summary),
  "member_hash": _compute_hash(member_summaries),
  "updated_at": datetime.now(timezone.utc).isoformat(),
  },

backend/tests/integration/test_encryption_roundtrip.py

@@ -46,6 +46,13 @@
  "Upload stuck"),
  ("issue_reports.description", "issue_reports", "id", "rich-issue-1",
  "description", "Syllabus upload spins forever."),
+ ("flashcards.front", "flashcards", "id", "rich-fc-cs-1", "front",
+ "What is a variable?"),
+ ("flashcards.back", "flashcards", "id", "rich-fc-cs-1", "back",
+ "A named storage location for a value."),
+ ("room_summaries.summary", "room_summaries", "room_id",
+ "rich-room-study-group", "summary",
+ "The group is reviewing recursion before the midterm."),
 ]
 
 # (label, id_value, column, expected_number)
@@ -123,6 +130,14 @@ def test_quiz_context_context_json_is_ciphertext_and_decrypts(db_conn):
  assert decoded["asked"] == 2
 
 
+def test_study_guides_content_is_ciphertext_and_decrypts(db_conn):
+ """#518: study_guides.content needs the JSON pair (encrypt_json/decrypt_json)."""
+ raw = _raw(db_conn, "study_guides", "id", "rich-guide-cs-f25-mid", "content")
+ assert isinstance(raw, str), "content stored as PLAINTEXT JSONB — encryption regressed"
+ decoded = decrypt_json(raw)
+ assert decoded["exam"] == "Midterm Exam"
+
+
 def test_documents_concept_notes_uses_json_encryption(db_conn):
  """documents.concept_notes is stored via encrypt_json (a list of concepts)."""
  raw = _raw(db_conn, "documents", "id", "rich-doc-cs-syllabus", "concept_notes")

backend/tests/test_flashcard_import_service.py

@@ -59,6 +59,28 @@ def test_filters_by_topic_when_offering_id_is_none(self):
  assert "offering_id" not in str(call_kwargs)
  assert "course_id" not in str(call_kwargs)
 
+ def test_dedup_matches_against_encrypted_existing_cards(self):
+ # #518: dedupe must be decrypt-aware — existing rows are ciphertext.
+ from services.encryption import encrypt
+ with patch("services.flashcard_import_service.table") as t:
+ t.return_value.select.return_value = [
+ {"front": encrypt("What is a variable?")}
+ ]
+ new = [{"front": "What is a variable?", "back": "A named storage location."}]
+ keep, skipped = svc.dedup_against_existing("u1", "c1", new)
+ assert keep == []
+ assert skipped == new
+
+ def test_dedup_still_matches_legacy_plaintext_rows(self):
+ # Pre-backfill rows are still plaintext — decrypt_if_present falls
+ # back to the raw value, so dedupe keeps working unchanged.
+ with patch("services.flashcard_import_service.table") as t:
+ t.return_value.select.return_value = [{"front": "What is a variable?"}]
+ new = [{"front": "What is a variable?", "back": "A named storage location."}]
+ keep, skipped = svc.dedup_against_existing("u1", "c1", new)
+ assert keep == []
+ assert skipped == new
+
 
 # ── check_rate_limit ─────────────────────────────────────────────────────────
 

CLAUDE.md

@@ -106,7 +106,7 @@ make explore # Chapter 2: bounded AI exploration of the ru
 
 ## Gotchas
 
-- Column-level encryption is on for sensitive columns: `user_profiles.name`/`first_name`/`last_name`/`bio`/`location` (these moved off `users` to `user_profiles` in the 0024 identity split), Google OAuth tokens, `messages.content`, `room_messages.text`, `sessions.summary_json`, `documents.summary` + `concept_notes` + `extracted_text` (the RAG OCR text added in 0030), `notes.title`/`body`/`last_summary`, `assignments.notes`/`points_possible`/`points_earned` (the enrollment-keyed gradebook table; points columns carry numeric semantics — use `decrypt_numeric` at read), `feedback.comment`/`topic` + `issue_reports.topic`/`description` (free-text user input, #520), and `quiz_attempts.questions_json`/`answers_json` + `quiz_context.context_json` (quiz performance data, #521; scalar analytics columns — score/total/difficulty/completed_at — stay plaintext). Helpers live in `backend/services/encryption.py`; use `encrypt_if_present` at write boundaries and `decrypt_if_present` / `decrypt_numeric` at read boundaries (including before injecting into AI prompts). `ENCRYPTION_KEY` must be set (32 bytes as 64 hex chars; generate via `python -c "import secrets; print(secrets.token_hex(32))"`). Deliberate exception: `newsletter_emails.email` stays plaintext (ADR 0026) — the UNIQUE constraint, lookup index, and both subscribe/allowlist upserts key on the value, and AES-GCM's per-call nonce breaks value equality.
+- Column-level encryption is on for sensitive columns: `user_profiles.name`/`first_name`/`last_name`/`bio`/`location` (these moved off `users` to `user_profiles` in the 0024 identity split), Google OAuth tokens, `messages.content`, `room_messages.text`, `sessions.summary_json`, `documents.summary` + `concept_notes` + `extracted_text` (the RAG OCR text added in 0030), `notes.title`/`body`/`last_summary`, `assignments.notes`/`points_possible`/`points_earned` (the enrollment-keyed gradebook table; points columns carry numeric semantics — use `decrypt_numeric` at read), `feedback.comment`/`topic` + `issue_reports.topic`/`description` (free-text user input, #520), `quiz_attempts.questions_json`/`answers_json` + `quiz_context.context_json` (quiz performance data, #521; scalar analytics columns — score/total/difficulty/completed_at — stay plaintext), and `flashcards.front`/`back` + `study_guides.content` + `room_summaries.summary` (derived content, #518; `study_guides.content` uses the JSON pair, `room_summaries.summary` keys its cache on the separate plaintext `member_hash` column so encryption doesn't affect cache-hit lookups). Helpers live in `backend/services/encryption.py`; use `encrypt_if_present` at write boundaries and `decrypt_if_present` / `decrypt_numeric` at read boundaries (including before injecting into AI prompts). `ENCRYPTION_KEY` must be set (32 bytes as 64 hex chars; generate via `python -c "import secrets; print(secrets.token_hex(32))"`). Deliberate exception: `newsletter_emails.email` stays plaintext (ADR 0026) — the UNIQUE constraint, lookup index, and both subscribe/allowlist upserts key on the value, and AES-GCM's per-call nonce breaks value equality.
 - Knowledge-graph mastery is now an append-only `node_mastery_events` table (replaced the `graph_nodes.mastery_events` JSON column in 0023); node/edge dedup is enforced by UNIQUE constraints. Don't read/write a `mastery_events` column.
 - Optional cross-worker cache (#97): `services/cache.py` wraps Redis and is **off by default** — with no `REDIS_URL` set it's a zero-overhead no-op and never fails a request (any Redis error → clean miss + warning). Currently backs the content-addressed OCR/extraction cache (`extraction_service.extract_text_from_file`, keyed on `sha256(file_bytes)` + engine). The `redis` dependency is only imported when `REDIS_URL` is set.
 - HTTP caching (#99): conditional GETs use `services/http_cache.py` (`make_etag`/`conditional`/`cached_json`). `Cache-Control` on these routes is **always `private`, never `public`** — the responses carry user-scoped, app-decrypted columns that must never be cached at a shared proxy/CDN. Derive the ETag from cheap change-keys (ids, `updated_at`, existing content hashes), not from the fully-built payload.

backend/db/backfill_encryption.py

@@ -302,6 +302,18 @@ def backfill_quiz_context(apply: bool) -> dict:
  return _encrypt_json_column("quiz_context", "context_json", pk="id", apply=apply)
 
 
+def backfill_flashcards(apply: bool) -> dict:
+ return _encrypt_text_column("flashcards", ["front", "back"], pk="id", apply=apply)
+
+
+def backfill_study_guides(apply: bool) -> dict:
+ return _encrypt_json_column("study_guides", "content", pk="id", apply=apply)
+
+
+def backfill_room_summaries(apply: bool) -> dict:
+ return _encrypt_text_column("room_summaries", ["summary"], pk="room_id", apply=apply)
+
+
 RUNNERS: dict[str, Callable[[bool], dict]] = {
  "users": backfill_users,
  "user_settings": backfill_user_settings,
@@ -315,6 +327,9 @@ def backfill_quiz_context(apply: bool) -> dict:
  "issue_reports": backfill_issue_reports,
  "quiz_attempts": backfill_quiz_attempts,
  "quiz_context": backfill_quiz_context,
+ "flashcards": backfill_flashcards,
+ "study_guides": backfill_study_guides,
+ "room_summaries": backfill_room_summaries,
 }
 
 

backend/db/seed_local_rich.py

@@ -553,7 +553,7 @@ def seed_notes_documents() -> None:
  )
 
 
-# (fc_id, user_id, offering_id, topic, front, back) — plaintext, grouped by topic.
+# (fc_id, user_id, offering_id, topic, front, back) — 🔒 front/back, grouped by topic.
 _FLASHCARDS = [
  ("rich-fc-cs-1", USER_ACTIVE, OFF_CS_F25, "CS Basics",
  "What is a variable?", "A named storage location for a value."),
@@ -579,8 +579,9 @@ def seed_flashcards() -> None:
  "user_id": user_id,
  "offering_id": off_id,
  "topic": topic,
- "front": front,
- "back": back,
+ # 🔒 front / back (#518)
+ "front": encrypt_if_present(front),
+ "back": encrypt_if_present(back),
  },
  )
 
@@ -625,11 +626,26 @@ def seed_study_guides() -> None:
  "offering_id": off_id,
  "exam_id": exam_id,
  "generated_at": generated_at,
- "content": content,
+ # 🔒 content (#518)
+ "content": encrypt_json(content),
  },
  )
 
 
+def seed_room_summaries() -> None:
+ # #518: room_summaries.summary is 🔒. PK is room_id (no id column), so this
+ # can't go through insert_if_absent.
+ if not table("room_summaries").select("room_id", filters={"room_id": f"eq.{ROOM_STUDY}"}):
+ table("room_summaries").insert({
+ "room_id": ROOM_STUDY,
+ "summary": encrypt_if_present("The group is reviewing recursion before the midterm."),
+ "member_hash": "rich-member-hash-v1",
+ })
+ h.record("room_summaries", created=True)
+ else:
+ h.record("room_summaries", created=False)
+
+
 # (qa_id, concept_node_id, difficulty, score, total, questions_json, answers_json, completed_at)
 _QUIZ_ATTEMPTS = [
  ("rich-qa-cs-variables-1", "rich-node-cs-variables", "easy", 9, 10,
@@ -766,6 +782,7 @@ def seed_sessions() -> None:
  "schools", "courses", "course_offerings", "users", "user_profiles", "user_roles",
  "enrollments", "graph_nodes", "graph_edges", "node_mastery_events",
  "gradebook_categories", "assignments", "rooms", "room_members", "room_messages",
+ "room_summaries",
  "notes", "documents", "flashcards", "study_guides", "quiz_attempts", "quiz_context",
  "sessions", "messages", "feedback", "issue_reports",
 ]
@@ -785,6 +802,7 @@ def main() -> None:
  seed_notes_documents()
  seed_flashcards()
  seed_study_guides()
+ seed_room_summaries()
  seed_quiz()
  seed_feedback()
  seed_sessions()

backend/e2e_oracles/gather.py

@@ -195,6 +195,10 @@ def run_counts(args: argparse.Namespace) -> tuple[list[Finding], int]:
  ("quiz_attempts", "id", "questions_json"),
  ("quiz_attempts", "id", "answers_json"),
  ("quiz_context", "id", "context_json"),
+ ("flashcards", "id", "front"),
+ ("flashcards", "id", "back"),
+ ("study_guides", "id", "content"),
+ ("room_summaries", "room_id", "summary"),
 )
 
 

backend/routes/flashcards.py

@@ -14,7 +14,7 @@
 from services.academics import resolve_offering, term_id_for_label
 from services.auth_guard import require_self, get_session_user_id
 from services.achievement_service import check_achievements
-from services.encryption import decrypt_if_present, decrypt_json
+from services.encryption import decrypt_if_present, decrypt_json, encrypt_if_present
 from services.flashcard_import_service import (
  dedup_against_existing,
  check_rate_limit,
@@ -236,8 +236,8 @@ def generate(body: GenerateFlashcardsBody, request: Request):
  "id": str(uuid.uuid4()),
  "user_id": body.user_id,
  "topic": body.topic,
- "front": c["front"],
- "back": c["back"],
+ "front": encrypt_if_present(c["front"]),
+ "back": encrypt_if_present(c["back"]),
  "times_reviewed": 0,
  "last_reviewed_at": None,
  "created_at": now,
@@ -261,8 +261,16 @@ def generate(body: GenerateFlashcardsBody, request: Request):
  except Exception:
  pass
 
+ # rows_to_insert holds the ciphertext just written; the response the
+ # frontend renders (freshly generated cards, before any list re-fetch)
+ # must carry plaintext front/back, not the encrypted insert payload.
+ response_cards = [
+ {**row, "front": decrypt_if_present(row["front"]), "back": decrypt_if_present(row["back"])}
+ for row in rows_to_insert
+ ]
+
  return {
- "flashcards": rows_to_insert,
+ "flashcards": response_cards,
  "context_used": {
  "documents_found": len(documents),
  "weak_concepts_found": len(weak_concepts),
@@ -291,6 +299,9 @@ def get_flashcards(
  "id,user_id,topic,offering_id,front,back,times_reviewed,last_rating,last_reviewed_at,created_at",
  filters=filters, order="created_at.desc"
  ) or []
+ for r in rows:
+ r["front"] = decrypt_if_present(r.get("front"))
+ r["back"] = decrypt_if_present(r.get("back"))
  if semester:
  # Term scoping (#141). This route is user-wide (no course id), so
  # the filter works on the cards' offering: cards from the selected
@@ -397,8 +408,8 @@ def import_commit(body: ImportCommitBody, request: Request):
  "user_id": body.user_id,
  "topic": body.topic,
  "offering_id": offering_id,
- "front": c["front"],
- "back": c["back"],
+ "front": encrypt_if_present(c["front"]),
+ "back": encrypt_if_present(c["back"]),
  "times_reviewed": 0,
  "last_reviewed_at": None,
  "created_at": now,

backend/routes/study_guide.py

@@ -4,6 +4,7 @@
 Study guide generation and caching.
 """
 
+import logging
 import uuid
 from datetime import datetime, timezone
 
@@ -24,10 +25,17 @@
 )
 from services.graph_service import get_courses as graph_get_courses
 from services.auth_guard import require_self
-from services.encryption import decrypt_if_present, decrypt_json
+from services.encryption import (
+ decrypt_if_present,
+ decrypt_json,
+ decrypt_json_column,
+ encrypt_json,
+)
 from services.http_cache import cached_json, conditional, make_etag
 from services.request_context import current_request_id
 
+logger = logging.getLogger(__name__)
+
 router = APIRouter()
 
 
@@ -198,7 +206,7 @@ def _generate_and_insert(user_id: str, offering_id: str, exam_id: str) -> dict:
  "offering_id": offering_id,
  "exam_id": exam_id,
  "generated_at": now,
- "content": content,
+ "content": encrypt_json(content),
  }
  table("study_guides").insert(row)
 
@@ -245,7 +253,14 @@ def get_cached_guides(user_id: str, request: Request):
 
  result = []
  for g in guides:
- content = g.get("content") or {}
+ try:
+ content = decrypt_json_column(g.get("content")) or {}
+ except Exception:
+ logger.warning(
+ "get_cached_guides: content decrypt failed for guide %s; degrading",
+ g.get("id"),
+ )
+ content = {}
  course_id = offering_to_course.get(g.get("offering_id"))
  result.append({
  "id": g["id"],
@@ -330,7 +345,7 @@ def get_guide(
  )
  if cached:
  row = cached[0]
- return {"guide": row["content"], "generated_at": row["generated_at"], "cached": True}
+ return {"guide": decrypt_json_column(row["content"]), "generated_at": row["generated_at"], "cached": True}
 
  result = _generate_and_insert(user_id, offering_id, exam_id)
  return {"guide": result["content"], "generated_at": result["generated_at"], "cached": False}

backend/services/flashcard_import_service.py

@@ -33,6 +33,7 @@
 from agents.flashcard import flashcard_agent
 from agents.usage import record_agent_usage
 from services import extraction_service
+from services.encryption import decrypt_if_present
 
 logger = logging.getLogger(__name__)
 
@@ -76,7 +77,9 @@ def dedup_against_existing(
  filters["topic"] = f"eq.{topic}"
 
  existing = table("flashcards").select("front", filters=filters) or []
- existing_norm = [_normalize(r.get("front", "")) for r in existing]
+ existing_norm = [
+ _normalize(decrypt_if_present(r.get("front", "")) or "") for r in existing
+ ]
 
  keep: list[Card] = []
  skipped: list[Card] = []

backend/services/social_cache_service.py

@@ -12,6 +12,7 @@
 from datetime import datetime, timezone
 
 from db.connection import table
+from services.encryption import decrypt_if_present, encrypt_if_present
 
 
 def _compute_hash(member_summaries: list[str]) -> str:
@@ -26,15 +27,15 @@ def get_cached_summary(room_id: str, member_summaries: list[str]) -> str | None:
  filters={"room_id": f"eq.{room_id}"},
  )
  if rows and rows[0]["member_hash"] == current_hash:
- return rows[0]["summary"]
+ return decrypt_if_present(rows[0]["summary"])
  return None
 
 
 def save_summary(room_id: str, member_summaries: list[str], summary: str) -> None:
  table("room_summaries").upsert(
  {
  "room_id": room_id,
- "summary": summary,
+ "summary": encrypt_if_present(summary),
  "member_hash": _compute_hash(member_summaries),
  "updated_at": datetime.now(timezone.utc).isoformat(),
  },

backend/tests/integration/test_encryption_roundtrip.py

@@ -46,6 +46,13 @@
  "Upload stuck"),
  ("issue_reports.description", "issue_reports", "id", "rich-issue-1",
  "description", "Syllabus upload spins forever."),
+ ("flashcards.front", "flashcards", "id", "rich-fc-cs-1", "front",
+ "What is a variable?"),
+ ("flashcards.back", "flashcards", "id", "rich-fc-cs-1", "back",
+ "A named storage location for a value."),
+ ("room_summaries.summary", "room_summaries", "room_id",
+ "rich-room-study-group", "summary",
+ "The group is reviewing recursion before the midterm."),
 ]
 
 # (label, id_value, column, expected_number)
@@ -123,6 +130,14 @@ def test_quiz_context_context_json_is_ciphertext_and_decrypts(db_conn):
  assert decoded["asked"] == 2
 
 
+def test_study_guides_content_is_ciphertext_and_decrypts(db_conn):
+ """#518: study_guides.content needs the JSON pair (encrypt_json/decrypt_json)."""
+ raw = _raw(db_conn, "study_guides", "id", "rich-guide-cs-f25-mid", "content")
+ assert isinstance(raw, str), "content stored as PLAINTEXT JSONB — encryption regressed"
+ decoded = decrypt_json(raw)
+ assert decoded["exam"] == "Midterm Exam"
+
+
 def test_documents_concept_notes_uses_json_encryption(db_conn):
  """documents.concept_notes is stored via encrypt_json (a list of concepts)."""
  raw = _raw(db_conn, "documents", "id", "rich-doc-cs-syllabus", "concept_notes")

backend/tests/test_flashcard_import_service.py

@@ -59,6 +59,28 @@ def test_filters_by_topic_when_offering_id_is_none(self):
  assert "offering_id" not in str(call_kwargs)
  assert "course_id" not in str(call_kwargs)
 
+ def test_dedup_matches_against_encrypted_existing_cards(self):
+ # #518: dedupe must be decrypt-aware — existing rows are ciphertext.
+ from services.encryption import encrypt
+ with patch("services.flashcard_import_service.table") as t:
+ t.return_value.select.return_value = [
+ {"front": encrypt("What is a variable?")}
+ ]
+ new = [{"front": "What is a variable?", "back": "A named storage location."}]
+ keep, skipped = svc.dedup_against_existing("u1", "c1", new)
+ assert keep == []
+ assert skipped == new
+
+ def test_dedup_still_matches_legacy_plaintext_rows(self):
+ # Pre-backfill rows are still plaintext — decrypt_if_present falls
+ # back to the raw value, so dedupe keeps working unchanged.
+ with patch("services.flashcard_import_service.table") as t:
+ t.return_value.select.return_value = [{"front": "What is a variable?"}]
+ new = [{"front": "What is a variable?", "back": "A named storage location."}]
+ keep, skipped = svc.dedup_against_existing("u1", "c1", new)
+ assert keep == []
+ assert skipped == new
+
 
 # ── check_rate_limit ─────────────────────────────────────────────────────────