docs: add overtone shift analysis and insights

- Add deep dive analysis on SVD axis overtone shift (docs/research/) - Update brainstorm with results documenting key finding: stability and overtone shift are independent phenomena - Add learning doc about axis stability vs semantic drift independence - Update drift report with detailed findings
2026-04-05 19:40:23 +02:00
parent dafdfd5370
commit 67eda93cb1
4 changed files with 1353 additions and 5 deletions
@@ -0,0 +1,79 @@
+---
+title: "SVD Axis Stability and Overtone Shift are Independent Phenomena"
+date: 2026-04-05
+module: analysis
+problem_type: insight
+component: motion-analysis
+severity: medium
+tags: [svd, overtone-shift, semantic-drift, axis-stability, parliamentary-analysis]
+applies_when:
+  - Interpreting SVD axes over multiple time windows
+  - Comparing motion content across different parliamentary periods
+  - Understanding why stable axis labels don't guarantee stable motion content
+---
+
+# SVD Axis Stability and Overtone Shift are Independent Phenomena
+
+## Key Insight
+
+When analyzing SVD axes across time windows, **axis stability** and **overtone shift** measure fundamentally different phenomena:
+
+| Phenomenon | What it Measures | How to Compute |
+|------------|-----------------|----------------|
+| **Axis Stability** | Whether the same motions/embeddings load on an axis | Lasso regression: `SVD_score ~ fused_embedding`, compare weight vectors via cosine similarity + Jaccard |
+| **Overtone Shift** | How motion content evolves over time | Semantic gravity (weighted mean embedding) tracking via cosine distance |
+
+**The implication**: An axis can be "stable" (parties vote similarly across years) while its semantic content drifts dramatically (different motions define the axis).
+
+## Evidence
+
+Analysis of 9 annual windows (2016-2026) revealed:
+
+- **9 of 10 axes are stable** (similarity > 0.7)
+- **All stable axes show high overtone shift** (1.3-1.97 cosine distance)
+- **No inflection points detected** — drift is gradual, not sudden
+
+### Example: Axis 1 (Coalition vs Opposition)
+
+| Metric | Value | Interpretation |
+|--------|-------|----------------|
+| Axis Stability | 0.83 | Coalition/opposition voting pattern is structurally consistent |
+| Overtone Shift | 1.47 avg, 1.97 max | Motion content defining "coalition" vs "opposition" has changed substantially |
+
+This means: PVV, VVD, NSC, BBB consistently vote together against SP, GL-PvdA, PvdD across all windows — but the specific motions that exemplify "coalition policy" in 2026 are semantically different from 2016.
+
+## Why This Matters
+
+1. **Axis labels are temporally bounded** — "Rechts kabinetsbeleid versus links oppositiebeleid" accurately describes 2016-2026, but the underlying motions have evolved.
+
+2. **Cross-temporal comparison is valid structurally but not semantically** — Party positions are comparable; motion content is not.
+
+3. **Axis 6 (Migration/Culture)** is an exception — Low stability (0.35) suggests fundamental change in how this dimension is structured.
+
+## How to Analyze This
+
+Use `scripts/motion_drift.py`:
+
+```bash
+uv run python scripts/motion_drift.py --db data/motions.db --output reports/drift
+```
+
+The script computes:
+- **Axis stability**: Lasso regression weights compared across windows
+- **Overtone shift**: Semantic gravity tracking
+- **Inflection points**: Sudden drift detection
+- **Party trajectories**: How parties move along stable axes
+
+## Prevention
+
+When updating SVD themes:
+1. Run `scripts/motion_drift.py` to check current overtone shift levels
+2. Verify that theme descriptions match current motion content, not historical content
+3. Monitor Axis 6 specifically for stability issues
+4. Consider temporal weighting in visualizations — recent windows better represent current semantics
+
+## Related
+
+- `scripts/motion_drift.py` — Analysis script
+- `docs/research/2026-04-05-svd-overtone-shift-deep-dive.md` — Deep analysis
+- `reports/drift/report.md` — Generated report