What changed
The 2026 Political Compass has had a substantial update. The short list:
- 57 candidates now plotted with real compass positions, up from a handful of hand-curated entries.
- Force-directed spacing so dots no longer pile into an unreadable line when candidates share similar positions.
- Corrected quadrant colours — Authoritarian Left is red, Authoritarian Right is blue, Libertarian Right is yellow, matching the standard political compass convention.
- AI-generated portraits for every candidate, described from actual photographs rather than generic prompts.
- Research-backed scoring — each candidate's compass position is now informed by policy statements, news coverage, interview transcripts, and published manifestos rather than a bare-bones bio.
The election date has also been corrected to 24 September 2026.
The candidate intelligence pipeline
This is the part that's worth explaining, because it's unusual.
Most election tools score candidates from a single source — a manifesto PDF, or a questionnaire the candidate fills in, or the editorial team's best guess. We wanted to build something that pulls from every public source available, cross-references them, and lets the AI scoring see the full picture before it assigns a position.
The pipeline runs in stages, each one feeding the next:
Stage 1: Discovery
An automated scraper walks the official candidate register and local news sites, building the candidate list. It captures names, constituencies, campaign links, and manifesto URLs. New candidates are added to the dataset; existing ones get their links enriched. Nothing is ever deleted — discovery is additive only.
This stage also handles name normalisation (accented characters, variant spellings, title prefixes like "Dr") and constituency name mapping, so the same candidate isn't accidentally duplicated under two slightly different spellings.
Stage 2: Research ingestion
A separate research pipeline reads structured dossiers that have been compiled for each candidate. Each dossier contains:
- Stated policy positions — what the candidate has said publicly, sourced and attributed.
- News coverage — relevant reporting on the candidate's public record, voting history, and policy announcements.
- Interview transcripts and media appearances — where available.
- Published manifestos — extracted as text from PDF documents.
This is the richest source of evidence. A candidate who has published a forty-page manifesto and given three interviews gets a much more confident compass score than one who has only announced their candidacy with a one-paragraph statement. That asymmetry is inherent to the source material, and we don't try to hide it — the compass page shows how many policy positions were analysed for each candidate.
Stage 3: AI enrichment
With the evidence assembled, each candidate is scored by Azure OpenAI. The model reads the full evidence package — bio, policy statements, news coverage, manifesto text — and returns:
- A political compass position (economic -10 to +10, social -10 to +10).
- Policy stances across ten categories (economy, healthcare, housing, education, environment, transport, digital, immigration, cost of living, governance), each with a score, a short label, and an evidence-linked explanation.
- A manifesto summary in structured HTML.
- Questions for the candidate — specific, data-informed questions that journalists or voters should ask, drawn from gaps, contradictions, or unsubstantiated claims in the evidence.
The model is prompted to be critical. If a candidate promises something they don't have the power to deliver, the scoring flags it. If statistics are cited without evidence, they're challenged. If positions contradict the candidate's track record, the inconsistency is highlighted.
Stage 4: Portrait generation
Each candidate's photograph is analysed by AI vision, which produces a detailed physical description — hair, glasses, clothing, expression. That description is then used to generate a flat-design geometric avatar in a consistent style across all candidates. The result is a set of portraits that are recognisably the right person but rendered as simple, clean icons rather than photographs.
This is a deliberate design choice. We don't host candidate photographs (rights, consent, and update logistics), but we want the compass and candidate cards to have faces rather than initials. The flat-design style keeps things neutral — no candidate looks more professional or approachable than another because of photograph quality.
Stage 5: Page data compilation
The final stage merges everything — candidate data, compass scores, policy positions, portraits, constituency boundaries, voter registration numbers, timeline events — into a single precomputed JSON file. The web page reads this file directly. There are no per-request AI calls when you load the compass; everything was computed ahead of time.
The spacing problem
The previous compass used a linear nudging algorithm to separate overlapping dots. When two candidates were close together, it pushed them apart along a single axis. This worked for small numbers but produced visible lines when many candidates clustered in the same region — which they do, because most Isle of Man candidates are centrist independents.
The new version uses force-directed repulsion. Every pair of dots that overlap pushes each other apart in both X and Y, iterating sixty times until the layout stabilises. Dots that land at exactly the same position (which happens when multiple candidates have no evidence and default to the centre) get a deterministic angular offset so they spread into a circle rather than sitting on top of each other. The result looks like a natural scatter rather than a ruled line.
What's next
The pipeline is designed to run repeatedly as new evidence appears. As candidates publish manifestos, give interviews, and make policy announcements in the run-up to September, the research dossiers can be updated and the enrichment re-run. Compass positions will shift as more evidence arrives — a candidate who was scored at (0, 0) due to lack of evidence will move to their real position once their manifesto is analysed.
We're also looking at a constituency comparison view and a voter-matching tool. More on those when they're ready.
- 2026 Political Compass — the updated page.
- Historical Compass — the same treatment applied to 2011, 2016, and 2021 elections.
- Election Hub — the full candidate directory.