diff --git a/thoughts/shared/designs/2026-03-28-compass-ui-improvements-design.md b/thoughts/shared/designs/2026-03-28-compass-ui-improvements-design.md
new file mode 100644
index 0000000..a7c9dbd
--- /dev/null
+++ b/thoughts/shared/designs/2026-03-28-compass-ui-improvements-design.md
@@ -0,0 +1,96 @@
+---
+date: 2026-03-28
+topic: "Compass UI Improvements"
+status: validated
+---
+
+# Compass UI Improvements
+
+## Problem Statement
+
+Three separate issues degrade the political compass UI:
+
+1. **SVD axis descriptions (axes 3–5) are outdated.** The current `label`, `explanation`, `positive_pole`, and `negative_pole` strings in `SVD_THEMES` were written for an earlier dataset and no longer match the structural patterns the axes actually capture. Using single-year (2024) centroid snapshots to verify this was misleading — the fix must be grounded in multi-year averages and motion-level content.
+
+2. **Y-axis direction indicators are broken.** The current `"Progressief ↑ / Conservatief ↓"` string is passed as the Plotly Y-axis title. Plotly rotates axis titles 90° counter-clockwise, so the arrows end up pointing sideways instead of up/down. This appears in the two compass scatter plots and in the trajectories tab.
+
+3. **No voting discipline context.** The compass shows where parties sit spatially but gives no sense of whether a party votes as a bloc. This context would make the compass more interpretable.
+
+## Constraints
+
+- No new DB tables or schema changes.
+- `compute_party_discipline` reads `mp_votes` where `mp_name LIKE '%,%'` (individual MP rows only — party-aggregate rows are excluded).
+- Skip discipline section if fewer than 5 roll-call motions in the selected window.
+- Follow existing patterns in `explorer.py` and `database.py`.
+- Tests use `uv run pytest`.
+
+## Approach
+
+**Change 1 — Axis descriptions:** Derive corrected descriptions from multi-year party centroid averages (all annual windows, not just 2024) and from the motion-level content that loads high/low on each axis. Update only `label`, `explanation`, `positive_pole`, `negative_pole` in `SVD_THEMES` entries for axes 3, 4, and 5. The `flip` boolean is not changed.
+
+**Change 2 — Y-axis arrows:** Replace the ↑/↓ characters from the axis title string (set to plain `"Progressief / Conservatief"`). Add two `fig.add_annotation` calls per chart: top-center `"▲ Progressief"` and bottom-center `"▼ Conservatief"`, using `xref="paper", yref="paper"`, `showarrow=False`, styled to be subtle (small font, muted color). Apply to both compass scatter plots and the trajectories chart.
+
+**Change 3 — Voting discipline:** Add a `compute_party_discipline(db_path, start_date, end_date)` function in `explorer.py` that queries individual MP votes, computes per-party Rice index (fraction voting with party majority), and returns a DataFrame with columns `[party, n_motions, discipline]`. In `build_compass_tab`, after rendering the compass chart, call this function with the window's date range, and render: (a) a horizontal bar chart sorted ascending (least disciplined at top), and (b) a small table showing the three most and three least disciplined parties. If fewer than 5 motions, show a brief explanatory message instead.
+
+## Architecture
+
+All changes are confined to `explorer.py`. No changes to `analysis/`, `database.py`, or test files (the discipline function is a read-only helper, not shared infrastructure).
+
+## Components
+
+**`SVD_THEMES` dict (explorer.py ~line 1156)**
+- Entries for axes 3, 4, 5 updated in-place.
+- New text is based on multi-year patterns (see Data Flow below).
+
+**Y-axis annotation helper (explorer.py)**
+- Small inline helper or inline code block that adds the two direction annotations to any given `fig`.
+- Called once after each `px.scatter(...)` and once after the trajectories `fig` is built.
+
+**`compute_party_discipline(db_path, start_date, end_date)` (explorer.py)**
+- Connects to DuckDB read-only.
+- Queries `mp_votes` filtered to individual MPs (`mp_name LIKE '%,%'`) and date range.
+- Groups by `(motion_id, party)`, counts votes per token, determines majority token, computes Rice index per motion per party.
+- Averages Rice index across motions per party.
+- Returns `pd.DataFrame(columns=["party", "n_motions", "discipline"])` or empty DataFrame.
+
+**`build_compass_tab` additions (explorer.py ~line 841+)**
+- After `st.plotly_chart(fig, ...)`, map the current `window_idx` to a `(start_date, end_date)` range.
+- Call `compute_party_discipline(...)`.
+- If result has ≥ 5 motions: render bar chart + extremes table under a `st.subheader("Stemgedrag cohesie")`.
+- If not: `st.caption("Te weinig hoofdelijke stemmingen voor cohesieanalyse.")`.
+
+## Data Flow
+
+**Axis description research (prior to implementation):**
+Multi-year centroid averages are computed by averaging each party's SVD vector across all annual windows in which it appears. The axis 3/4/5 descriptions are updated to reflect these stable patterns rather than any single year's snapshot.
+
+**Discipline computation:**
+```
+mp_votes (individual MPs, date range)
+  → GROUP BY (motion_id, party, vote) → vote counts
+  → determine majority_vote per (motion_id, party)
+  → Rice index = (count voting with majority) / (total voting) per motion per party
+  → average Rice index across motions → per-party score
+  → return DataFrame
+```
+
+## Error Handling
+
+- `compute_party_discipline` returns an empty DataFrame on any DB exception (logged, not raised), following the pattern of other read helpers in `explorer.py`.
+- Empty DataFrame → show the "too few motions" caption (same path as < 5 motions).
+- The Y-axis annotation is purely visual — no error paths needed.
+- Axis description changes are static strings — no runtime risk.
+
+## Testing Strategy
+
+- The `compute_party_discipline` function is tested with a small in-memory DuckDB fixture in `tests/test_political_compass.py`:
+  - Construct a fixture with 6 motions, 2 parties, each with varying vote splits.
+  - Assert returned DataFrame has correct columns and that discipline scores are in [0, 1].
+  - Assert empty DataFrame is returned when date range has 0 motions.
+- Y-axis annotation: no unit test needed (visual only, trivially correct).
+- Axis description changes: no unit test needed (static strings).
+- Run all tests with `uv run pytest tests/test_political_compass.py -v` after each change.
+
+## Open Questions
+
+None. All design decisions are resolved.