Method and sources

How prlmnt works

Before an election, parties publish manifestos with hundreds of pledges. What parliament decides afterwards is public record – but spread across thousands of entries, transcripts and PDF documents. prlmnt brings these official sources together, so you can see how parties vote and how that matches what they announced.

prlmnt doesn’t recommend whom to vote for and doesn’t rate political positions. We apply the same methods to every party, show what the sources say and link every statement to its source, so it can be checked.

Our sources

• Austrian Parliament – Open Data. All motions, government bills and committee reports of the National Council (Nationalrat) in the current legislative period (since 24 October 2024), with voting results, topics and speeches.
• Parliamentary Correspondence (Parlamentskorrespondenz). Reports by parliament’s own press service on plenary and committee sessions since January 2024 – including the positions of the parliamentary groups and statements by members.
• Manifestos for the 2024 general election by ÖVP, SPÖ, FPÖ, NEOS and the Greens, as published by the parties as PDF files. The texts are the parties’ own work; we quote from them and refer to the page in the original.

Parliament’s data is licensed under CC BY 4.0. We update it daily. All sources are in German; we show titles and quotes in the original, so nothing is lost in translation.

Coming soon: press releases of the parliamentary groups, full speech transcripts from the stenographic records, and earlier legislative periods.

How we analyse votes

In Austria’s National Council, members usually vote along party lines. For every vote, parliament publishes which parliamentary groups voted in favour and which against. We extract this for every vote – including amendments and resolutions that never become law.

The agreement between two parties is the share of plenary votes in which both voted the same way. In detail:

• A bill is voted on in second and third reading, usually with the same result. We only count the third reading, so bills don’t count twice.
• When parts of a bill are voted on separately and a party supports only some of them, we record that as “partly” and leave it out of the agreement figures.
• Committee votes are not included: committees in Austria meet behind closed doors, and parliament usually publishes only their result, not how each party voted.
• “Adopted” refers to the individual vote, not always to the cause behind it. If the National Council adopts a committee report recommending rejection, for example, the original motion has failed.

How we analyse manifestos

1. Capture the text. We split every manifesto into headings and paragraphs and keep the page number, so every passage can be looked up in the original.
2. Check against the printed page. Text taken from PDF files can be wrong, for example because of badly encoded fonts or text that is stored but not visible. So we also read every page with optical character recognition (OCR) and compare each sentence with the printed text. Depending on the manifesto, 97 to over 99 percent of sentences are confirmed this way; the rest – mostly headings in decorative fonts or text on coloured backgrounds – is flagged.
3. Assign topics. A language model assigns each paragraph to up to three of the 20 policy areas parliament itself uses for motions and laws. This lets us connect manifestos and votes through the same topics later on.
4. Find and back up pledges. The model looks for concrete, verifiable pledges: what legislators, government or administration should do. Values, descriptions of the current situation and criticism of other parties don’t count. For every pledge it has to quote the passage word for word; we automatically check that the quote really appears in the manifesto and discard pledges without a matching quote.
5. Human review. The pledges found are suggestions. Only once a person has reviewed a pledge do we use it to compare with voting behaviour. This review is still to come.

Which language model we use

All five manifestos were analysed with the same model, so differences between parties don’t come from the tool. We chose it by comparing five models on the same 40 paragraphs, with the same instructions and the same checks. What counted was how often a model invents quotes, skips paragraphs or merges several pledges into one. The most reliable was Claude Haiku 5.5 by Anthropic: no invented quotes, no skipped paragraphs.

Language models still make mistakes. That’s why we show the original quote and its page next to every pledge.

How topic shares are calculated

The manifesto overview shows what share of each manifesto’s text goes to which topic. Every paragraph counts with its length; if it has several topics, it is split evenly between them. This way a manifesto’s shares add up to 100 percent and can be compared between manifestos, however long they are or however they are divided into paragraphs.

What the data doesn’t show

• Motives. A party may vote against a motion that matches its manifesto – out of coalition discipline, because it tabled its own motion, or because the text doesn’t go far enough for it.
• Individual members. Roll-call votes are rare in the National Council. We analyse how the parliamentary groups voted.
• Weight. Every vote counts the same, whether it’s the budget or a procedural question. Distinguishing types of votes is in progress.
• Importance. Topic shares measure how much text a manifesto devotes to a topic – not how important it is to the party or how concrete its proposals are.
• Number of pledges. How many pledges a manifesto contains depends strongly on its writing style. The number is not a basis for comparing parties.

What’s next

• Human review of the pledges found.
• A reading view of the manifestos with pledges highlighted and filterable by topic.
• The positions of the parliamentary groups from the Parliamentary Correspondence, and linking pledges to votes on the same topic.
• More countries.