Findings

01 The cross-axis structure

Of the 120 unique axis pairs, the 5 strongest correlations cluster into two groups: an Islamist–secular cluster (religion-in-state ↔ Palestinian-question ↔ regime-stance) and a regime-stance ↔ democratic-norms cluster. The most independent pairs run orthogonal to these — economic / federalism / regional-stance carry distinct explanatory power. Pairs are limited to those sharing at least 30 scored entities, so a thin scoped axis can't manufacture a headline bond.

Strongest correlations

Most independent

These pairs do distinct explanatory work — knowing a party's score on one tells you little about its score on the other.

02 Where parties cluster vs. where they spread

The standard deviation of scores on each axis reveals where MENA parties genuinely disagree (high σ) and where they reach near-consensus (low σ). Compass-axis cleavages run wide; some issue axes reach near-consensus across the region. These rankings run over the 12 core axes — the ones scored on at least 30 entities, enough to read a field-level pattern from. The 4 sparser scoped axes are held out into their own table below rather than averaged in: with a handful of points one outlier swings σ and "most contested" stops meaning contested. Why.

Widest spread (most contested)

Field spread (σ across parties) with the panel's mean within-cell disagreement beside it. An axis that is wide and rater-contested is unsettled on two counts; one that is wide but tightly rated is a real, agreed cleavage.

Tightest spread (consensus)

Most-skewed mean (regional consensus toward one pole)

Within the core, Pan-Arab vs particularist, Regime stance are well-covered but rater-contested (panel σ above 2.3): enough entities to rank, but the panel splits on the cell scores, so the composite there is an estimate, not a fixed point.

Peripheral axes — reported, not aggregated

These scoped axes are scored on fewer than 30 entities — too few to anchor a field-level claim — so they sit out of the rankings above and the headline correlations. They're shown in full, panel disagreement included, because the sparsity is the point: it marks where coverage is thin, not where the region agrees.

Distributions across the 16 axes

Each row shows the histogram of party scores binned at integer intervals on a fixed [-10, +10] scale. Blue bars = negative scores (left pole), red bars = positive (right pole), sage = zero (centre). The shape of each distribution makes the spread / consensus / skew patterns visible at a glance.

• Economic Statist → Market n=83 · μ=-0.7
• Social Authority → Libertarian n=83 · μ=-1.4
• State & religion Religious state → Secular state n=83 · μ=+0.1
• Liberal democracy Weak/anti → Strong commitment n=83 · μ=+0.3
• West alignment Anti-Western → Pro-Western n=81 · μ=+0.6
• Regional stance Resistance/maximalist → Stability/normalization n=82 · μ=-0.3
• Palestinian question Opposed → Pro-Palestinian rights n=83 · μ=+5.7
• Civil liberties Restrict → Expand n=81 · μ=-0.0
• Regime stance anti-regime → pro-regime n=78 · μ=+2.1
• Pan-Arab vs particularist particularist → pan-Arab n=66 · μ=+0.8
• Centralism vs federalism centralist → federalist n=78 · μ=-3.6
• Traditionalism vs modernization traditionalist → modernizing n=78 · μ=+1.2
• Gender equalityscoped Patriarchal traditionalism → Gender equality n=22 · μ=-3.0
• Iran posturescoped Anti-Iran / adversarial → Pro-Iran / aligned n=19 · μ=-1.6
• Press freedomscoped State-controlled press → Free press n=21 · μ=-2.5
• Sectarian power-sharingscoped Consociational / quota system → Post-sectarian / civic state n=11 · μ=-4.5

03 Structural status of MENA party politics

The government_role distribution captures something compass scores can't: the operational political space MENA parties live in. 44 of the 98 parties on file (45%) are banned or extra-parliamentary. Only 10 are lead parties.

• 10 lead parties — the dominant force in their country's government
• 16 in coalitions (major / minor / confidence-supply)
• 28 in opposition
• 39 extra-parliamentary — legal but no seats
• 5 banned / outlawed

Banned, dissolved, restricted

5 outlawed parties currently on file (1 in IQ; 1 in DZ; 2 in IL; 1 in SY); 8 dissolved; 0 restricted. Inclusion of outlawed entities is deliberate — for many MENA contexts, the most consequential political actors are also the ones the state is trying to suppress (Northern Islamic Movement in IL, FIS in DZ, FJP in EG, Al-Wefaq in BH).

04 Where parties concentrate

The top 5 countries by party count carry the bulk of the dataset's analytical weight. Coverage in the others is editorial: a chosen anchor set, not a comprehensive parliamentary roster.

• 27 parties — IL
• 9 parties — EG
• 7 parties — TN
• 7 parties — PS
• 6 parties — LB

05 Per-axis poles

The party at each end of every axis — the lowest (most negative) score and the highest (most positive). Useful for calibration: if either pole party feels meaningfully wrong at its stated extreme, the axis rubric or the scoring deserves another look. Note that on some axes both poles sit on the same side of zero (e.g. democracy, where even the most pro-democracy party scores well below the most anti-democracy one).

06 Declared vs. behavioral — where rhetoric and record split

For 20 parties across 8 axes, the dataset carries both a declared score (manifesto-stated position) and a behavioral score (voting record / governance behaviour). The gap between them is the measure of how much a party's rhetoric overstates or understates its actual record. Across all 36 pair rows, the mean absolute spread is 6.08 on the −10 to +10 scale (max 12), and 35 of 36 point one direction: declared is more positive than behavioral — the rhetoric overstates the record more often than understates it.

Largest divergences

Top 8 (party, axis) pairs by absolute gap. Negative spread (red) means the party's record runs to the left of its rhetoric on that axis; positive (green) means the record runs further right.

Party	Axis	Declared	Behavioral	Δ
Hezbollah	Sectarian power-sharing	+3.0	-9.0	+12.0
Free Patriotic Movement	Sectarian power-sharing	+2.0	-9.0	+11.0
Hamas	Liberal democracy	+4.0	-7.0	+11.0
Ansar Allah (Houthis)	Liberal democracy	+3.0	-8.0	+11.0
Mostaqbal Watan	Liberal democracy	+3.0	-7.0	+10.0
National Liberation Front	Liberal democracy	+5.0	-5.0	+10.0
Hezbollah	Liberal democracy	+3.0	-5.0	+8.0
Freedom and Justice Party	Liberal democracy	+6.0	-2.0	+8.0

Toggle the lens chips at the top of the compass to see how it moves between declared and behavioral views for these parties.

07 Inter-rater reliability — the reproducibility pilot

The panel has layers that are easy to conflate, so to be explicit: author-baseline is the hand-coded prior; r1 (Gemini, Grok, no definitions) is a locked "before", kept only to measure whether the rubric changed the scores, and excluded from the published median; r2 (nine models — Claude, Gemini, GPT, Gemma, Grok, Kimi, DeepSeek, Qwen, MiniMax — with the definitions and the declared/behavioral rule) is the live multi-model panel, landing country by country and shown on /ratings; a grounded pilot (one model scoring from cited documents) seeds the document-verification round. The κ below is an earlier, separate coder pilot. The formal multi-model κ over r2 lands once coverage is broad enough to be representative — a two-country panel would mislead.

The dataset's hand-coded composite scores are the Pilot Phase baseline. To check whether those scores are reproducible from the published rubrics rather than idiosyncratic to one coder, a stratified sample of (party, axis) cells was re-coded across 4 rounds by 2 independent second-pass coders, blind to the baseline scores. Quadratic-weighted Cohen's κ and Pearson r measure agreement; mean |Δ| is the average gap on the −10..+10 scale.

This is an internal reproducibility check, not external expert validation — both coders share the project's rubric framework, so the absolute agreement runs high and should be read as a floor, not proof. The signal worth acting on is the relative pattern across axes. External multi-coder validation by independent domain experts is a v0.3 milestone.

Headline: 102 second-pass ratings over 51 cells · 7 axes · overall κ 0.90 , Pearson r 0.91 , mean |Δ| 1.77 points.

Pairwise agreement

Each second-pass coder against the baseline, plus the two coders against each other.

Agreement by axis — hardest first

Baseline vs. pooled second-pass ratings. The governance axes carry the most coder disagreement.

08 What's next

These are the patterns the current data shows. The next methodological step shifts how scores get assigned — from author-hand-coded against secondary literature to derived from reading party platforms, speeches, and voting records, with cited passages. The interesting result, when that lands: where the two methods agree on a party's position, and where they pull apart. More on methodology.

+ Inter-method agreement

For every (party, axis) where both a hand-coded composite score and a document-grounded declared/behavioral score exist, the pair tells us whether the two methods agree. Mean absolute difference (lower is better) and Pearson r (higher is better) per axis.

Overall: 19 parties have both a hand-coded and a document-grounded score on at least one axis. Mean absolute difference across all paired (party × axis) cells: 7.05 (on the −10 to +10 scale).

Mean |Δ| under 2 means the two methods are within one rubric anchor of each other — functionally agreeing. r above 0.6 means they rank parties in similar order on the axis even when absolute scores differ. Axes where both numbers are weak are the interesting cases: the rubric language probably needs work, or one of the methods is missing context the other has.