Why this site exists
A multi-AI analysis driven by the goal of avoiding partisanship.
This site is an experiment: can multiple AI models and strict editorial guardrails produce a programme analysis that remains useful to readers with opposite opinions? Every artefact (sources, prompts, raw outputs) is public. The reader is invited to verify, not to trust.
How an analysis is produced
Six steps, each linked to a public artefact. Following the chain lets you trace any claim back to the programme sentence that grounds it.
- 01
1. Consolidate sources
Raw sources (PDFs, web pages, transcripts) are consolidated into a single sources.md document, reviewed and validated by a human before any analysis runs.
- 02
2. Analyse (× N models)
The same versioned prompt is sent in parallel to several AI models. Each model produces an independent structured JSON, stored as-is in raw-outputs/.
- 03
3. Aggregate
A meta-model aggregates per-model outputs into a single structured aggregated.json, preserving disagreements in an agreement_map. Positioning scores are never averaged.
- 04
4. Human review
A CLI opens every contested or flagged claim. As long as one item remains unresolved by a human, the publish step is blocked.
- 05
5. Translate (optional)
Optional. A dedicated prompt translates only the prose fields listed in an allowlist, leaving structures and numbers untouched. The result is reviewed before ingestion.
- 06
6. Publish
Once human review is complete, the version is published by updating the current symlink. The site is rebuilt from the validated artefacts.
The five editorial guardrails
- 01
Analysis, not advocacy
The site describes tradeoffs. The verdict belongs to the reader.
A programme that widens the deficit is described as “increases the budget gap by €X bn/year”, not as “wrecks the public finances”.
- 02
Symmetric scrutiny
Each candidate is analysed on the same dimensions, with the same prompt, in the same order.
If a critical note is applied to programme A on its debt trajectory, the same grid is applied to every other programme — no exception, no softening.
- 03
Measurement, not indictment
Where a quantity exists, it is named. Adjectives do not replace numbers.
An intergenerational debt is expressed as “€X/person/year transferred from cohort A to cohort B”, not as “theft from the young”.
- 04
Dissent preserved
When models diverge, the disagreement is preserved in the structure, not erased by an average.
If three models place a programme at +2 on the economic axis and two place it at +4, the gap is displayed as-is — not reduced to an average of +2.8.
- 05
Radical transparency
Sources, prompts, raw outputs and code are public. Every artefact used to produce a claim is inspectable.
For any claim visible on the site, the originating programme sentence and the raw JSON produced by each model are reachable in two clicks.
Political positioning: ordinal, not cardinal
Each model places the programme on five axes (economic, societal, institutional, environmental, international) using integers from −5 to +5. These scores are ordinal: they indicate a rank relative to fixed historical anchor figures, not a measurable distance. For that reason, the site never computes an arithmetic mean between models; it displays the modal value, the consensus interval, and each individual position.
How models are aggregated
Model outputs are merged by a designated meta-model that produces a structured aggregate. Every claim carries an agreement_map listing the models that support it and those that contest it. No positioning score is arithmetic-mean averaged. A human review validates or rejects each aggregation before publication.
Analytical dimensions
Each programme is analysed on the same fixed set of dimensions, in the same order, with the same prompt. If a dimension is not addressed by a programme, “not addressed” is itself the finding — the dimension is never dropped for that candidate.
- Economic & fiscal
- Social & demographic
- Environmental & long-term
- Institutional & democratic
- Security & sovereignty
- Health
- Education
Transparency: where to verify
Everything is public. Here are the entry points to verify any claim made on the site.
What this site is not
- —Not a voting guide.
- —Not an endorsement platform for a candidate or party.
- —Not a fact-checking service for debate statements or social-media posts.
- —Not a policy-preference aggregator telling you which candidate matches your opinions.
- —Not a “neutral” platform that refuses to state findings — findings are stated clearly, without overlaid editorial.
- —Not a service funded by a party, a media outlet, a foundation or a sponsor.
Known limitations
Model limitations. AI models share part of their training data, pretraining biases, and a strong English-corpus skew. Aggregating several models reduces individual variance; it does not eliminate correlated error.
Source limitations. The pipeline reads what each candidate has published. An eloquent programme whose figures do not add up grades differently from a sparse programme: this finding reflects the programme, not the person.
Human limit. A single person reviews every sources.md and every aggregated.json. Project throughput is bounded by that review, and the bus-factor risk is acknowledged.
Governance and funding
- Maintainer: a single person. Side project run on spare time.
- Funding: none. No sponsor, no donations, no grants, no media partnership.
- Political affiliation: none declared. The code, prompts and aggregates are public and verifiable.
- Marginal cost: LLM API calls are paid by the maintainer, and most often executed via the manual or Copilot modes to limit expense.
- Open source code. Anyone can audit the methodology or reproduce it for other elections.