feat(extremity): two-dimensional rescoring with subagent pipeline

- 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

reports/overton_window/2d_extremity_correlation_report.md

@@ -0,0 +1,112 @@
+# Two-Dimensional Extremity Correlation Report
+
+**Date:** 2026-05-24
+**Motions scored:** 117 (stratified sample: ~25 per original extremity bucket)
+**Scoring model:** Deepseek v4 flash (subagents via project skill)
+
+## Purpose
+
+The original extremity score is a single 1–5 rating of policy radicalism. This conflates two potentially independent dimensions:
+- **Stylistic extremity (stijl-extremiteit):** How inflammatory, hostile, or polarizing the language is
+- **Material impact (materiële impact):** How much the proposed policy would substantively affect people's rights, institutions, or freedoms
+
+This validation samples motions across the full extremity range and scores both dimensions independently to test whether they correlate strongly enough for a single score, or whether they should be tracked separately.
+
+---
+
+## Results
+
+### Overall correlation
+
+| Metric | Value |
+|--------|-------|
+| N | 117 |
+| Pearson r | **0.453** (moderate) |
+| Mean stylistic | 2.01 |
+| Mean material | 2.86 |
+| Mean absolute difference | 1.11 |
+| S ≤ 2 AND M ≥ 3 (masking) | 43 (36.8%) |
+
+**r = 0.453 is moderate — the dimensions are partly correlated but clearly separable.** Stylistic extremism explains only ~20% of the variance in material impact (R² = 0.205). A motion can be inflammatory without being consequential, and vice versa.
+
+### Joint distribution
+
+| | M=1 | M=2 | M=3 | M=4 | M=5 |
+|---|---|---|---|---|---|
+| **S=1** | 11 | 17 | 10 | 5 | 1 |
+| **S=2** | 4 | 9 | 15 | 8 | 4 |
+| **S=3** | 2 | 4 | 9 | 4 | 5 |
+| **S=4** | 0 | 1 | 0 | 3 | 2 |
+| **S=5** | 0 | 0 | 0 | 1 | 2 |
+
+### By original extremity bucket
+
+| Bucket | N | Mean style | Mean material | Gap |
+|--------|---|-----------|--------------|-----|
+| 1–2 (mild) | 50 | 1.56 | 2.24 | +0.68 |
+| 2–3 (moderate) | 25 | 2.00 | 2.88 | +0.88 |
+| 3–4 (high) | 25 | 2.56 | 3.56 | +1.00 |
+| 4–5 (extreme) | 17 | 2.53 | 3.65 | +1.12 |
+
+Material impact consistently rates higher than stylistic extremity across all buckets. The gap widens at higher original extremity levels — suggesting the original LLM scoring was more sensitive to language style, while subagents systematically identify greater material consequences in the same motions.
+
+---
+
+## Key findings
+
+### 1. "Low style, high impact" is the dominant divergence pattern
+
+**36.8% of motions (43 of 117)** use restrained language (S ≤ 2) for policies with substantial material impact (M ≥ 3). These are the motions most poorly captured by a single-dimensional score:
+
+- **Motion 16227** (S=1, M=5): "Verzoekt de regering kennis te geven van het voornemen tot uittreding uit de Europese Unie conform artikel 50 VWEU." Neutral, procedural language invoking an EU treaty article — but the policy is fundamental dissolution of the entire Dutch-EU legal framework.
+
+- **Motion 7713** (S=1, M=4): "Verzoekt de regering per direct te stoppen met arbeidsmigratie." Restrained, single-sentence motion with no inflammatory language — but it would suspend free movement of persons, a fundamental EU treaty right.
+
+- **Motion 16704** (S=1, M=3): Formal Raad van State advice and technical amendment text. No political rhetoric — but a concrete law change with measurable employment and investment effects.
+
+- **Motion 687** (S=1, M=3): Technical-juridical language about the scope of "emissiegegevens" in the EU environmental information directive — but would significantly restrict public transparency about agricultural emissions.
+
+### 2. Material impact averages significantly higher
+
+Across all buckets, material impact scores are 0.68–1.12 points higher than stylistic scores. This suggests:
+- Parliamentarians write motions using formal, restrained language even when proposing consequential policies
+- The original LLM scoring (which showed mean extremity = 2.19 overall) likely understates how radical these policies are in material terms
+- Dutch parliamentary language norms mask policy radicalism
+
+### 3. "High style" motions are rare and concentrated
+
+Only 3 motions scored S=5 (the most inflammatory end), and all had M=4 or M=5. Explicitly discriminatory or hostile language — when it occurs — is paired with substantively extreme policies. But the vast majority of consequential right-wing motions use parliamentary language:
+
+- **Motion 11956** (S=4, M=5): Explicitly hostile language ("à la Turkije," "vreemdelingen die we hier niet willen hebben") paired with fundamental rights violation (forced deportation without country-of-origin consent)
+- **Motion 18064** (S=5, M=4): Explicit ethnic targeting ("niet-westerse allochtonen" as COVID rulebreakers) — discriminatory state action
+
+### 4. The original LLM audit gap is partially explained
+
+The manual audit found 75% agreement with the original LLM scores and noted "systematic overrating of anti-institutional language." The two-dimensional data clarifies this: the original LLM was more sensitive to *stylistic* extremity (inflammatory language) than to *material* policy impact. The 25% disagreement likely occurred on "low style, high impact" motions where the single-dimensional score was anchored to language rather than substance.
+
+---
+
+## Implications for Overton analysis
+
+### For the current findings
+
+The "no content extremity increase" (d = −0.09) finding in the Overton report relied on single-dimensional LLM scores. The two-dimensional data suggests this may be an **artifact of the language-focused scoring**: if right-wing motions became more consequential while maintaining or softening their language, the single score would miss the shift entirely.
+
+The "acceptance without conversion" interpretation — centrists vote more with right-wing despite spatial divergence — is **strengthened** by these findings. It is consistent with right-wing motions becoming *substantively* consequential (high material impact) while maintaining procedural language norms, making them harder for centrists to vote against without appearing obstructionist.
+
+### Recommendations
+
+1. **Re-score all 2,986 motions with two-dimensional scoring.** The moderate r = 0.453 confirms the dimensions are separable. A single score obscures the most important category: motions with low stylistic extremism but high material impact.
+
+2. **Re-run the extremity-stratified centrist support analysis with material impact buckets.** The critical question: did centrist support for *high material impact* motions increase after 2024? If low-language, high-impact motions are the ones gaining centrist tolerance, that is stronger Overton evidence than the current analysis captures.
+
+3. **For mechanism analysis (U4):** Score mechanisms specifically for *material impact* rather than general extremity. The question is not "how extreme is this motion?" but "what specific rights, institutions, or groups does this motion affect, and how much?"
+
+---
+
+## Data
+
+- **Full results:** `data/motions.db` → `extremity_scores_2d` (117 rows)
+- **Raw JSON:** `/tmp/extremity_2d_results.json`
+- **Scoring skill:** `.opencode/skills/score-extremity/SKILL.md`
+- **Orchestrator:** `analysis/right_wing/extremity_rescore_2d.py`