Add GL-PvdA merger SVD analysis design with findings

Investigation of GroenLinks-PvdA merger dynamics in SVD space:
- Finding 1: GL-PvdA were 2.8-10.5% of avg inter-party distance apart pre-merger
- Finding 2: Merged party started most cohesive (#1 in 2023) but now 55% above avg spread
- Finding 3: Converged to 4.5% by Q3 2023, essentially indistinguishable
- Finding 4: GL/PvdA were most stable parties (10-25% drift) while VVD/D66 moved 70-177%

thoughts/shared/designs/2026-04-16-glpvda-merger-svd-analysis-design.md

@@ -0,0 +1,113 @@
+---
+date: 2026-04-16
+topic: "GroenLinks-PvdA Merger Dynamics in SVD Space"
+status: validated
+---
+
+## Problem Statement
+
+We need concrete, data-driven findings about the GroenLinks-PvdA merger for a blog post. Four questions:
+1. How similar were GL and PvdA in SVD space before the merger?
+2. How cohesive is the merged party compared to others?
+3. When did GL and PvdA converge in SVD space?
+4. Which parties shifted the most, and how do GL/PvdA compare?
+
+## Constraints
+
+- Investigation-only: query the database, report findings, no code changes
+- Database: `data/motions.db` (DuckDB), SVD vectors in `svd_vectors` table
+- Party affiliation: use `mp_votes.party` for historical labels (not `mp_metadata` which only tracks current party)
+- SVD vectors are at MP level (`entity_type='mp'`); party centroids must be computed from MP vectors
+- Cross-window scale differences require normalization (distances as fraction of avg inter-party distance)
+
+## Approach
+
+**Method**: Compute party centroids from MP-level SVD vectors, grouped by `mp_votes.party` for historical party labels. Normalize all distances as fractions of the average inter-party distance in each window to make cross-window comparisons meaningful.
+
+**Data source**: `svd_vectors` table (MP vectors per window) joined with `mp_votes` (historical party labels per vote, majority party per MP per year).
+
+**Key discovery**: `mp_metadata` only tracks current party — pre-merger GL and PvdA MPs who merged are now labeled "GroenLinks-PvdA". We must use `mp_votes.party` for historical accuracy.
+
+## Architecture
+
+N/A — this is a data investigation, not a system design.
+
+## Components
+
+### Finding 1: GL-PvdA Pre-Merger Similarity
+
+GL and PvdA were already remarkably close in SVD space well before the merger:
+
+| Year | GL↔PvdA Distance | As % of Avg Inter-Party Distance | Nearest Other Party | Distance to Nearest |
+|------|------------------|----------------------------------|---------------------|---------------------|
+| 2019 | 2.10 | 10.5% | PvdD | 9.6 |
+| 2020 | 2.23 | 5.0% | CU | 28.7 |
+| 2021 | 1.46 | 4.4% | FVD | 11.6 |
+| 2022 | 1.16 | 2.8% | FVD | 6.5 |
+
+The nearest non-PvdA party to GL was always 5-10x further away than PvdA itself. They converged over time — from 10.5% of average inter-party distance in 2019 down to 2.8% in 2022.
+
+### Finding 2: Post-Merger Cohesion
+
+| Year | GL-PvdA Spread | Avg Other Spread | Ratio | Cohesion Rank |
+|------|---------------|-------------------|-------|---------------|
+| 2023 | 1.50 | 19.95 | 0.08 | #1 most cohesive |
+| 2024 | 14.05 | 18.47 | 0.76 | Mid-pack |
+| 2025 | 28.09 | 18.09 | 1.55 | Below average |
+| Current | 43.30 | 28.05 | 1.54 | Below average |
+
+The merged party started as the most cohesive party in parliament (2023), but by 2025 its internal spread is 55% above average — the merger created a party that's internally more diverse than typical Dutch parties.
+
+### Finding 3: Merger Convergence Timeline
+
+| Window | GL↔PvdA Distance | Normalized Ratio |
+|--------|------------------|------------------|
+| 2019-Q3 | 0.98 | 25.5% |
+| 2020-Q1 | 1.38 | 18.6% |
+| 2021-Q1 | 1.58 | 19.3% |
+| 2022-Q3 | 0.86 | 9.4% |
+| 2023-Q1 | 0.58 | 7.1% |
+| **2023-Q3** | **0.37** | **4.5%** |
+| 2023-Q4 | 0.46 | 5.5% |
+
+By Q3 2023 — just before the formal merger — GL and PvdA centroids were only 4.5% of the average inter-party distance apart. Essentially indistinguishable in voting pattern space.
+
+### Finding 4: Large Positional Shifts
+
+GL and PvdA were the most stable parties in parliament (normalized drift per year):
+
+| Period | GL Drift | PvdA Drift | VVD Drift | D66 Drift | PVV Drift |
+|--------|----------|------------|-----------|-----------|-----------|
+| 2019→2020 | 14.5% | 16.6% | 140.2% | 145.6% | 121.9% |
+| 2020→2021 | 21.8% | 25.8% | 115.8% | 82.2% | 207.3% |
+| 2021→2022 | 11.6% | 10.8% | 70.8% | 91.2% | 51.9% |
+| 2022→2023 | 54.5% | 23.3% | 109.7% | 177.3% | 222.1% |
+
+While VVD and D66 moved 70-177% per year, GL and PvdA drifted only 10-25%. The merger partners were anchored in place while the rest of the landscape shifted.
+
+## Data Flow
+
+1. Query `svd_vectors` for MP vectors per window
+2. Join with `mp_votes` to determine each MP's majority party in that year
+3. Compute party centroids as mean of member vectors
+4. Compute pairwise distances and normalize by average inter-party distance
+5. Track convergence timeline using quarterly windows
+
+## Error Handling
+
+- Windows with insufficient MPs (<3 per party) are excluded from centroid calculations
+- The `mp_votes.party` column uses multiple label variants ("GroenLinks", "GL", "GroenLinks-PvdA") — normalized in queries
+- The 2023 transition year has mixed labels (some GL, some PvdA, some GL-PvdA) — handled by majority-vote assignment per MP
+
+## Testing Strategy
+
+N/A — data investigation. Key validation checks:
+- Cross-reference MP counts with known parliament compositions
+- Verify that GL + PvdA MP counts match expected seat counts per year
+- Confirm that convergence timeline aligns with known political events (merger announcement Oct 2023)
+
+## Open Questions
+
+- Should we compute cosine similarity instead of Euclidean distance for cross-window normalization?
+- The 2025 and current_parliament windows show very different absolute scales — should we normalize vectors before computing distances?
+- The few remaining "GL" (8) and "PvdA" (5) labeled MPs in 2025 may be artifacts — should they be included in the GL-PvdA group?