Skip old-schema grading templates so Stats cannot crash

README.md

@@ -174,7 +174,7 @@ See [`WORKFLOW.md`](WORKFLOW.md) for the full round-setup flow.
 | What | Where | Safe? |
 |------|--------|-------|
 | **en-us human scores (old 55-Q schema)** | `definitions/gradings/{MK,ML,RS}/*.json` in git — 80 terms × 3 graders (~4.8 MB) | **Yes — in repo** |
-| **sv-se / hr-hr human scores** | Were on HF dataset `TransLegal/grading-answers` (`{jurisdiction}/users/{username}_answers.parquet`); reports in `hf-spaces/gradings-analysis/` prove they existed (Mar–Apr 2026) | **Verify HF access** — not mirrored in `definitions/gradings/` |
+| **sv-se / hr-hr human scores** | Historical files on HF dataset `TransLegal/grading-answers` (old category/subcategory schema). Removed from the live tree; recover from dataset commit `4ba3d14f048c1b76aefb93e7051a42a9eeff0a32`. Reports in `hf-spaces/gradings-analysis/` | **Git history on the dataset** |
 | **Streamlit app answers (any jurisdiction)** | HF **bucket** `TransLegal/grading-answers` when deployed (`*_answers.csv` plus `user_progress.csv`); hourly copy on the HF **dataset**; local mode under `{GRADING_DATA_DIR}/{jurisdiction}/users/` | Local only if you graded locally |
 | **AI template rows** | `grading_template.parquet` — regeneratable from `structured_analysis` | Not precious |
 
@@ -227,7 +227,7 @@ Hourly dataset versioning runs from **this GitHub repo**, not from the Space. Ad
 1. **Jurisdiction Discovery:** Local mode scans `GRADING_DATA_DIR`; HF mode scans the dataset for `grading_template.parquet`
 2. **User Accounts:** Each jurisdiction has separate user accounts (same username can exist in different jurisdictions)
 3. **Data Storage:** Local mode writes parquet and `user_progress.csv` under `GRADING_DATA_DIR/{jurisdiction}/users/`; HF mode writes sparse CSV grades and the navigation table to the storage bucket and keeps templates / `users.json` on the dataset
-4. **Stats:** After sign-in, **Stats** in the sidebar opens a read-only dashboard of Pass / fail / ungraded cards for you, everyone, and each other grader, across every jurisdiction that has a template. Opening Stats does not change last-viewed / furthest-viewed. Local demo data: `uv run python seed_stats_demo.py` from this directory (new accounts use password `1234`; existing nonempty verdicts are kept).
+4. **Stats:** After sign-in, **Stats** in the sidebar opens a read-only dashboard of Pass / fail / ungraded cards for you, everyone, and each other grader, across every jurisdiction whose template matches the current swipe-deck schema. Older leftover templates are skipped (they do not crash the page). Opening Stats does not change last-viewed / furthest-viewed. Local demo data: `uv run python seed_stats_demo.py` from this directory (new accounts use password `1234`; existing nonempty verdicts are kept).
 
 ## Deployment
 

src/stats.py

@@ -17,6 +17,26 @@ DEFAULT_FAIL_REASONS: tuple[str, ...] = (
 )
 OTHER_REASON = "Other"
 _ANSWERS_SUFFIXES = ("_answers.parquet", "_answers.csv")
+SWIPE_TEMPLATE_COLUMNS: tuple[str, ...] = (
+    "term",
+    "field_of_law",
+    "concept",
+    "section",
+    "section_index",
+    "question_id",
+    "group",
+    "part",
+    "question",
+    "question_index",
+    "answer",
+    "source_fields",
+)
+
+
+def missing_template_columns(columns: object) -> tuple[str, ...]:
+    """Return swipe-deck columns absent from a template frame."""
+    present = {str(name) for name in columns}
+    return tuple(name for name in SWIPE_TEMPLATE_COLUMNS if name not in present)
 
 
 @dataclass(frozen=True)

src/streamlit_app.py

@@ -40,6 +40,7 @@ from stats import (
     counts_from_verdicts,
     dashboard_html,
     grades_to_verdicts,
+    missing_template_columns,
     overlay_verdicts,
     sum_counts,
     username_from_answers_filename,
@@ -290,65 +291,66 @@ def get_jurisdiction_display_mapping(jurisdiction_codes):
 # ---------------------------------------------------------------------------
 # Template + user data I/O
 # ---------------------------------------------------------------------------
-TEMPLATE_COLUMNS = (
-    "term",
-    "field_of_law",
-    "concept",
-    "section",
-    "section_index",
-    "question_id",
-    "group",
-    "part",
-    "question",
-    "question_index",
-    "answer",
-    "source_fields",
-)
+@st.cache_data
+def _read_template_df(jurisdiction: str) -> pd.DataFrame | None:
+    """Load a jurisdiction parquet with no schema check and no Streamlit stop."""
+    try:
+        if LOCAL_MODE:
+            template_path = _local_template_path(jurisdiction)
+            if not template_path.is_file():
+                return None
+            return pd.read_parquet(template_path)
+        file_path = hf_hub_download(
+            repo_id=HF_DATASET_REPO,
+            filename=f"{jurisdiction}/grading_template.parquet",
+            repo_type="dataset",
+            token=HF_TOKEN,
+        )
+        return pd.read_parquet(file_path)
+    except Exception:
+        return None
+
+
+@st.cache_data
+def discover_compatible_jurisdictions() -> tuple[str, ...]:
+    """Jurisdictions whose template matches the current swipe-deck columns."""
+    compatible: list[str] = []
+    for code in discover_available_jurisdictions():
+        frame = _read_template_df(code)
+        if frame is None:
+            continue
+        if missing_template_columns(frame.columns):
+            continue
+        compatible.append(code)
+    return tuple(compatible)
 
 
 @st.cache_data
 def load_grading_template(jurisdiction):
     """Load the PASS/FAIL question grading template (local path or HF)."""
-    available = discover_available_jurisdictions()
-    if not jurisdiction or jurisdiction not in available:
+    compatible = discover_compatible_jurisdictions()
+    if not jurisdiction or jurisdiction not in compatible:
+        available = discover_available_jurisdictions()
         st.error("❌ **Error**: Jurisdiction is not set or invalid.")
-        if available:
-            st.error(f"Available jurisdictions: {', '.join(available)}")
-        st.stop()
-
-    if LOCAL_MODE:
-        template_path = _local_template_path(jurisdiction)
-        try:
-            if not template_path.is_file():
-                raise FileNotFoundError(f"Template not found: {template_path}")
-            df = pd.read_parquet(template_path)
-        except Exception as e:
-            st.error(f"❌ **Error loading local grading template**: {str(e)}")
-            st.error(f"Expected file at `{template_path}`")
-            st.stop()
-    else:
-        try:
-            file_path = hf_hub_download(
-                repo_id=HF_DATASET_REPO,
-                filename=f"{jurisdiction}/grading_template.parquet",
-                repo_type="dataset",
-                token=HF_TOKEN,
-            )
-            df = pd.read_parquet(file_path)
-        except Exception as e:
+        if compatible:
+            st.error(f"Available jurisdictions: {', '.join(compatible)}")
+        elif available:
             st.error(
-                f"❌ **Error loading grading template from Hugging Face**: {str(e)}"
+                "Found template files, but none match the current grading schema. "
+                "Rebuild them with `build_grading_template.py`."
             )
-            st.error(
-                f"Expected `{jurisdiction}/grading_template.parquet` in {HF_DATASET_REPO}"
-            )
-            st.stop()
+        st.stop()
+
+    df = _read_template_df(jurisdiction)
+    if df is None:
+        st.error("❌ **Error**: Could not load the grading template.")
+        st.stop()
 
-    missing = [c for c in TEMPLATE_COLUMNS if c not in df.columns]
+    missing = missing_template_columns(df.columns)
     if missing:
         st.error(
             "❌ **Error**: grading template is missing columns "
-            f"{missing}. Rebuild it with `build_grading_template.py`."
+            f"{list(missing)}. Rebuild it with `build_grading_template.py`."
         )
         st.stop()
     return df
@@ -593,8 +595,13 @@ def _build_stats_snapshot(
 ]:
     per_user: dict[str, GradeCounts] = {}
     per_jurisdiction: dict[str, GradeCounts] = {}
+    used: list[str] = []
     for code in jurisdictions:
-        cards = get_cards(load_grading_template(code))
+        frame = _read_template_df(code)
+        if frame is None or missing_template_columns(frame.columns):
+            continue
+        used.append(code)
+        cards = get_cards(frame)
         total_cards = len(cards)
         for grader in list_grader_usernames(code):
             counts = _counts_for_user(
@@ -620,10 +627,10 @@ def _build_stats_snapshot(
         ((name, counts) for name, counts in per_user.items() if name != username),
         key=lambda item: (-item[1].graded, item[0]),
     )
-    names = get_jurisdiction_display_mapping(tuple(jurisdictions))
+    names = get_jurisdiction_display_mapping(tuple(used))
     jurisdiction_rows = [
         (code, names.get(code, code), per_jurisdiction.get(code, sum_counts()))
-        for code in jurisdictions
+        for code in used
     ]
     return everyone, you, others, jurisdiction_rows
 
@@ -657,6 +664,8 @@ def _render_stats_page(
     username: str, jurisdiction: str, jurisdictions: list[str]
 ) -> None:
     _render_stats_fabs()
+    compatible = set(discover_compatible_jurisdictions())
+    skipped = [code for code in jurisdictions if code not in compatible]
     everyone, you, others, jurisdiction_rows = _build_stats_snapshot(
         username,
         jurisdiction,
@@ -674,6 +683,12 @@ def _render_stats_page(
         ),
         unsafe_allow_html=True,
     )
+    if skipped:
+        st.caption(
+            "Skipped templates that are not the current grading schema: "
+            + ", ".join(skipped)
+            + "."
+        )
 
 
 def _progress_file_missing(exc: BaseException) -> bool:
@@ -1551,7 +1566,7 @@ def _try_register(username: str, password: str, jurisdiction: str) -> None:
 # Login page
 # ---------------------------------------------------------------------------
 if not st.session_state.logged_in:
-    available_jurisdictions = discover_available_jurisdictions()
+    available_jurisdictions = list(discover_compatible_jurisdictions())
     auto_user = os.getenv("GRADING_AUTO_LOGIN") if LOCAL_MODE else None
     if auto_user and available_jurisdictions:
         jurisdiction = st.session_state.jurisdiction or available_jurisdictions[0]
@@ -1621,7 +1636,7 @@ elif notice == "register":
 username = st.session_state.username
 jurisdiction = st.session_state.jurisdiction
 
-available_jurisdictions = discover_available_jurisdictions()
+available_jurisdictions = list(discover_compatible_jurisdictions())
 if not jurisdiction or jurisdiction not in available_jurisdictions:
     st.error("❌ **Error**: Jurisdiction is not set. Please sign out and in again.")
     if st.button("Sign out"):
@@ -1629,7 +1644,7 @@ if not jurisdiction or jurisdiction not in available_jurisdictions:
     st.stop()
 
 if st.session_state.view == "stats":
-    _render_stats_page(username, jurisdiction, available_jurisdictions)
+    _render_stats_page(username, jurisdiction, list(discover_available_jurisdictions()))
     st.stop()
 
 st.markdown('<div class="grading-top-align"></div>', unsafe_allow_html=True)

tests/test_stats.py

@@ -9,6 +9,7 @@ from stats import (
     dashboard_html,
     empty_counts,
     grades_to_verdicts,
+    missing_template_columns,
     overlay_verdicts,
     percent,
     segment_widths,
@@ -17,6 +18,40 @@ from stats import (
 )
 
 
+def test_missing_template_columns_detects_old_schema() -> None:
+    old = pd.DataFrame(
+        columns=[
+            "term",
+            "category",
+            "category_index",
+            "subcategory",
+            "subcategory_index",
+            "question",
+            "answer",
+        ]
+    )
+    missing = missing_template_columns(old.columns)
+    assert "field_of_law" in missing
+    assert "question_id" in missing
+    current = pd.DataFrame(
+        columns=[
+            "term",
+            "field_of_law",
+            "concept",
+            "section",
+            "section_index",
+            "question_id",
+            "group",
+            "part",
+            "question",
+            "question_index",
+            "answer",
+            "source_fields",
+        ]
+    )
+    assert missing_template_columns(current.columns) == ()
+
+
 def test_empty_user_is_all_ungraded() -> None:
     counts = counts_from_verdicts({}, 10)
     assert counts.passed == 0