- Project-local skill .opencode/skills/score-extremity/ for subagent dispatch
- Orchestrator extremity_rescore_2d.py with load_skill/sample/format/validate/store
- 16 TDD tests covering all orchestrator functions
- 117 motions scored by deepseek v4 flash subagents (12 parallel batches)
- Pearson r=0.45 between stylistic and material dimensions — separable
- Key finding: 36.8% of motions use restrained language for consequential policies
- 2d_extremity_correlation_report.md documents distribution, divergence patterns,
and implications for the Overton acceptance-without-conversion narrative
- Reclassified centrist to {D66, CDA, CU, NSC} — removing VVD/BBB
which are center-right coalition partners
- Added centrist_support_strict (0.251→0.507, d=+0.65), center_right_support,
and left_support_mp columns via migration script
- Figure 1 now shows center-right (VVD/BBB) support as orange dashed line
- New Figure 3: bar chart of left-party support for right-wing motions
(0.268→0.202, left opposition hardened)
- New report Section 6 covering left-wing support trends
- All analysis now uses strict centrist definition throughout
Extremity Scorer (U4 enhanced):
- Now scores BOTH original motion text AND layman explanation separately
- Schema: text_score, text_explanation, layman_score, layman_explanation
- Text scores: 1→7, 2→33, 3→5, 4→5 (mild-to-moderate)
- Layman scores: 1→12, 2→20, 3→17, 4→1 (slightly milder)
Sentiment Analysis (U5 enhanced):
- Now scores BOTH original motion text AND layman explanation separately
- Schema: text_score, text_explanation, layman_score, layman_explanation
- Text sentiment avg: 0.294 (slightly positive)
- Layman sentiment avg: 0.416 (more positive - summaries tone down hostility)
Category Derivation (new):
- Two-phase LLM approach: derive taxonomy from sample, then apply to all
- Discovered 7 categories from 30-motion sample:
veiligheid/justitie, corona/pandemie, economie/belasting, klimaat/milieu,
defensie/buitenland, asiel/vreemdelingen, overig
- Applied to 50 motions with distribution shown in DB
- Adds category + category_explanation columns to right_wing_motions
Implements U5: sentiment_analysis.py uses LLM batch calls (fallback when no
local Dutch sentiment model is available) to score motion sentiment on [-1, 1]
scale.
Design:
- Prompt asks for sentiment from -1 (hostile/aggressive) to 1 (constructive)
- JSON schema enforces numeric score + Dutch explanation
- Batch size 10, max_workers 5 for parallel API calls
- Stores results in table
- Updates with avg_sentiment, sentiment_std,
pct_strongly_negative per year
Sample validation (50 motions): good variance across [-0.9, 1.0] range.
Implements U3: temporal_analysis.py computes yearly_summary from the
right_wing_motions table (U2 output).
Metrics per year:
- total_right_wing, pct_of_total, total_motions
- avg_right_support, avg_left_opposition, centrist_support
- avg_right_keyword_matches, extremity_index (U4 placeholder)
- yoy_right_wing_delta, yoy_pct_delta
Key finding: right-wing motions grew from ~4% (2018) to ~12% (2024-2025)
of all motions, with rising centrist support over time.
Removes embedding-heavy search and generic browser tabs from the Explorer.
The project does not currently use embeddings meaningfully, so the similarity
search and browser features were dead weight.
Changes:
- explorer.py: Remove search and browser tabs, keep compass/trajectories/SVD
- explorer.py: Remove fused_embeddings and similarity_cache stats from sidebar
- Home.py: Update Explorer description to match new focused layout
- analysis/tabs/__init__.py: Remove search and browser exports
- tests: Update decomposition and import tests for new tab set
Result: Explorer now has 3 focused analytical tabs instead of 5.
- compute_svd_for_window now computes explained variance ratio (s²/sum(s²))
and appends it as a metadata row (entity_type='metadata',
entity_id='explained_variance') to motion_rows
- load_scree_data reads this metadata row from svd_vectors instead of
querying the non-existent sv_metadata column
- run_svd_for_window counts only entity_type='motion' rows in stored_motion
so metadata rows don't inflate the count
- Added 5 TDD tests covering load, compute, store, and round-trip
All 227 tests pass.
- load_scree_data: return [] with TODO until schema stores EVR metadata
- load_party_axis_scores: compute from vectors instead of missing table
- load_party_axis_scores_for_window: same vector-based fallback
- load_party_scores_all_windows[_aligned]: check table existence,
fall back to computing from load_positions when absent
All functions predated decomposition (5afbad1, 2026-04-05) and relied on
party_axis_scores / sv_metadata columns that were never created.
- Fix mindmodel-schedule.yml to use uv and Python 3.13
- Add pytest.yml for push/PR test gate
- Remove broken scheduler service from docker-compose.yml
- Consolidate config.py into analysis/config.py with backward-compat shim
- Rewrite README.md with quickstart and project overview
- Update pre-commit-config.yaml to enable black, ruff, isort hooks
- Add pyright type-check job (continue-on-error until baseline fixed)
- Update AGENTS.md with Gitea infrastructure note
- analysis/explorer_data.py: add AND window_id NOT LIKE '%-Q%' to
_UNIFORM_DIM_SQL so quarterly windows are filtered at the source
- explorer.py: remove stale comment justifying quarterly inclusion;
remove redundant '-Q' guard in SVD tab trajectory view
- scripts/recompute_svd.py: replace quarter_bounds() with year_bounds()
that handles annual window IDs like '2024'; filter window list to
annual-only before recomputing SVD
Revise SVD_THEMES labels based on TF-IDF analysis of top 50 motions
per component (pool size: current_parliament). Manual review of motion
titles ensures labels reflect actual parliamentary content rather than
party position semantics.
Key corrections:
- Axis 1: fiscal/economic policy vs social welfare + international rights
- Axis 4: active international engagement vs restraint
- Axis 5: pragmatic financial support vs progressive individual rights
- Axis 6: fossil fuels/financial incentives vs climate/intl rights
- Axis 7: practical-administrative vs idealistico-procedural (kept)
- Axis 8: European defense cooperation vs domestic socioeconomic policy
- Axis 9: concrete-administrative vs systemic reform
- Axis 10: citizen protection vs government regulation
Subagent analysis caught that axes 5 and 6 are NOT the same
(Nationale soevereiniteit) — manual motion review confirms distinct
content for each. Axes 1, 5, 6 had completely wrong labels.
Refs: thoughts/explorer/svd_label_review.md
See also: docs/brainstorms/2026-04-13-topic-derived-svd-labels-requirements.md
- Add compute_nd_axes() for N-component PCA with Procrustes alignment
- Add _get_aligned_party_scores() helper in explorer.py
- Update build_svd_components_tab to use aligned scores for all components
- Compute flip direction from aligned score centroids using CANONICAL_LEFT/RIGHT
Previously the compass (political_axis.py) used hardcoded party sets that
excluded Volt and PvdD, while the SVD components tab (svd_labels.py) used
CANONICAL_LEFT/RIGHT which includes them. This caused inconsistencies in
axis orientation where Volt appeared most left on the compass but PvdD
appeared most left in the SVD components visualization.
Changes:
- Import CANONICAL_LEFT/RIGHT from config in political_axis.py
- Replace hardcoded party sets with CANONICAL_LEFT/RIGHT for axis orientation
- Update tests to match new SVD_THEMES labels
Redo theme analysis after pool-based motion assignment change.
New labels reflect actual motion content per component:
1. Economische sectorbelangen versus sociale welvaart
2. Nationalistische versus multilateralistische oriëntatie
3. Verzorgingsstaat versus defensie en nationale veiligheid
4. Internationale instituties en multilateralisme versus nationale soevereiniteit
5. Gemeenschapszin versus individuele rechten
6. Ecologische transitie versus economische conservatie
7. Praktisch-bestuurlijk versus idealistisch-proceduraal
8. Internationale samenwerking versus nationale soevereiniteit
9. Pragmatische probleemoplossing versus regulering
10. Minder overheidsbemoeienis versus meer handhaving
Previously, components 1-2 in the SVD tab used Procrustes-aligned PCA
coordinates (from load_positions), which meant the SVD tab showed PCA
dimensions of the 50D aligned space rather than the actual raw SVD
components. This was a fundamental inconsistency — the SVD tab's component 2
showed completely different party ordering than the raw SVD component 2.
Changes:
- explorer.py: Unified all components 1-10 to use raw SVD values via
load_party_axis_scores_for_window(). Removed the separate
load_positions() path for components 1-2. Now all components use the
same data source (50D vectors from svd_vectors table).
- explorer.py: Updated flip computation to cover ALL components 1-10
(was range 3-11 for components 3-10 only). The compute_flip_direction
function correctly determines sign for each component.
- explorer.py: Unified rendering to always use _render_party_axis_chart_1d
(was _render_party_axis_chart for components 1-2 using 2D coords).
- explorer.py: Unified trajectory to always use load_party_scores_all_windows.
- analysis/config.py: Updated component 1 label (simplified explanation,
removed coalition-specific policy references).
- analysis/config.py: Updated component 2 label to "Nationalistisch versus
kosmopolitisch" matching raw SVD data (PVV/FVD at positive extreme,
Volt/DENK/GL-PvdA at negative extreme).
- tests: Updated test assertions to match new labels.
- scripts/validate_svd_themes.py: Verified all components pass right-wing
alignment check, config flip consistency, and theme pole consistency.
Fixes the core inconsistency: SVD tab component 2 now uses the same raw
SVD data as components 3-10, with consistent party ordering and labels.
The compass remains a separate PCA-based visualization.
Allow analysis modules to be imported in lightweight test environments
without duckdb installed. Modules that need duckdb for actual queries
still require it at runtime, but import-time failures are handled gracefully.
Two related bugs fixed:
1. Label alignment: Removed static left_pole/right_pole from SVD_THEMES
entries. These labels assumed a fixed flip direction but could mismatch
with runtime flip computation, causing right-wing parties to appear on
the wrong side. Labels are now always derived from positive_pole,
negative_pole, and the runtime flip direction.
2. Score mismatch: Changed tijdtraject view for components 3-10 from
load_party_scores_all_windows_aligned() to load_party_scores_all_windows().
Procrustes alignment rotates the full 50-dim vector space to align
components 1-2, but this also transforms components 3-10, making their
scores incomparable with the single-window view. Per-window flip
computation already handles orientation alignment for these components.
Also updated svd_labels.py to prefer analysis.config as the canonical
source for SVD_THEMES, falling back to explorer only when config is
unavailable.
- Add CANONICAL_RIGHT (PVV, FVD, JA21, SGP) and CANONICAL_LEFT frozensets
to analysis/config.py as the canonical source of truth
- Update analysis/svd_labels.py to import from config; re-export as
RIGHT_PARTIES/LEFT_PARTIES for backward compatibility
- Add build_window_party_scores helper to analysis/explorer_data.py
- Add 7 integration tests in tests/test_axis_political_orientation.py
validating that canonical right parties appear on the right side of SVD
axes (x=component 1, y=component 2) using real DuckDB data
Move rng initialization before the party loop so each party gets a
unique segment of the random stream instead of identical sequences.
Replace Python bootstrap loop with vectorized numpy indexing.
Pure numpy function that computes bootstrap confidence intervals for
party centroid vectors. Handles N>=2 (bootstrap), N=1 (degenerate CI),
and N=0 (excluded) cases. Uses np.random.default_rng for reproducibility.
Both _load_window_ids and _load_mp_vectors_for_window only read from the DB.
Opening without read_only=True caused an IOException when Streamlit already held
a read-only lock, silently returning an empty scree plot.
Previously load_scree_data computed L2-norms per dimension on current_parliament
vectors only, giving ~11% for PC1. This was inconsistent with the compass which
uses all windows + Procrustes alignment and gets PC1=24.1%.
Added compute_svd_spectrum() helper to political_axis.py that reuses the same
alignment pipeline. load_scree_data now delegates to it. _render_scree_plot
no longer re-normalizes (inputs are already EVR percentages). Hover label
updated to 'verklaarde variantie'.
Quarterly windows (29 of 41 total) diluted PC1 explained variance ratio
from ~20% down to ~14.6%. The fix splits the vector collection loop into:
- pca_vecs: annual windows only (re.match r'^\d{4}$') -> M_pca used for SVD
- all_vecs: every window -> M used for projections onto derived axes
Centering for SVD and global_mean for projection both now use M_pca.mean(axis=0)
so axes are consistent. Falls back to all windows if no annual windows exist.
