How we do it
29N Responde analyses the interviews and debates of the campaign for the 29 November election, applying the same rules to every candidate. It does not recommend how to vote, does not score anyone and does not run polls.
How we classify each answer
We split each interview into questions, including follow-up questions, and check whether the answer addresses what was asked: the figure, the deadline, the yes or the no, the position.
- Answers. Answers everything that is asked: the figure, the deadline, the yes or the no.
- Partly answers. Answers part of it. We indicate what was left unanswered.
- Does not answer the question. Talks about something else or declines to answer. We show the question and the answer so you can check.
Rules we always apply:
- The question is always visible: not all candidates get the same questions or the same format.
- We assess the whole sequence: if the answer comes after a follow-up question, it counts.
- Rebutting the premise of a question with a specific argument is not “not answering”.
- Saying “yes” or “no” to a yes-or-no question counts as answering, even without further detail.
- We do not assess whether the answer is true, good or convincing: only whether it addresses what was asked.
- We also measure the interviewer: how much they speak, how often they interrupt, how many follow-up questions they ask.
- We do not add anything up into an overall score, nor do we rank the candidates.
We also measure how long it takes for the first sentence addressing the question to come. It is a data point, not a judgement: some answers need context.
Claims and fact-checks
We flag factual claims (figures, comparisons, facts, cause-and-effect relationships, attributions to others), whether they can be checked with public data and where. If a figure is given without saying where it comes from, we point it out: “figure without a source”.
- Our checks only say whether the official data supports or contradicts the claim, with the figure, the source and the date; or that it cannot be verified with the available data.
- We do not use “lie”, “false” or “demagoguery”: they describe intentions or judgements, not data.
- When an independent fact-checker has already checked a claim, we link to their work with their own label and name.
- Every check is reviewed by a person before it is published.
Commitments and manifestos
Only what the candidate says they, their party or their government will do counts as a commitment (“we will”, “we are going to”, “I will not”). “We need to”, “we want”, “we propose”, what they would do in a hypothetical case and descriptions of a proposal do not count. That is why the number of commitments depends on how each candidate speaks and does not measure how many proposals they have: that is what their manifesto is for.
When a candidate promises something in an interview, we check whether they make it specific (figure, deadline, mechanism, funding) and look for it in their party's manifesto:
- Matches the manifesto, with a link to the proposal and its page.
- More specific than the manifesto: gives a figure or a deadline that the manifesto does not.
- Not found in the manifesto: this does not mean it contradicts it.
- Differs from the manifesto: only with both quotes side by side and after human review.
- The party has not yet published a manifesto, where that is the case.
If a commitment gives no figure, deadline, mechanism or funding, we mark it as an “unspecified promise”.
How we use artificial intelligence
- The transcript comes from the broadcast's own subtitles or from an automatic transcription of the audio. Who is speaking at each moment is proposed by the AI from the text, and a person always checks it against the video.
- A language model (Claude Opus 5.5, by Anthropic) splits the interview into questions and proposes the classification. For each interview we say which model did the analysis; when a second model classifies it separately with the same instructions, the differences go to human review.
- The model does not write the timestamps: it indicates which sentence it is quoting and the program calculates the exact second from the transcript.
- Every quote is checked automatically: if it does not appear word for word at that point in the video, it is not published.
- Always the same model for everyone. If the model refuses to analyse something, no other model is used: it is reviewed by hand.
- We publish the exact instructions: AI instructions.
The AI assistant
- When you ask it a question, our server searches the content of this website (the proposals in the 2026 manifestos, the analysed interviews and the party profiles) for the most closely related excerpts, spread across the parties so that none of them dominates the answer.
- It sends those excerpts and your question to Claude Opus 5.5 (Anthropic), with public instructions: answer only from those excerpts, cite each one, do not recommend how to vote, do not rank or score the parties, do not discuss polls and present the parties in the fixed order used across the website.
- If something is not in the excerpts, it says so. It can make mistakes: every statement links to its source so you can check it.
- We do not store your questions or the answers. It does not answer during the day of reflection or on polling day.
- Its exact instructions: AI instructions.
Human review
There is a single reviewer. To minimise the influence of their own view, they review blind: without names or parties, in a queue that mixes interviews and candidates. They always review:
- Who each voice is attributed to.
- All answers classified as “Does not answer the question”.
- All cases where the two models disagree.
- A random sample of the rest.
Anything that both models classify the same way as “Answers” or “Partly answers”, and that is not in the sample, is published without individual review. Before publishing, we measure the agreement between the model and human review on a hand-annotated interview (Cohen's kappa, minimum 0.6): [pending].
Review by party
| Party | Questions published | Reviewed by a person | Corrected |
|---|---|---|---|
| PP | 153 | 0 | 0 |
| PSOE | 202 | 0 | 0 |
| Vox | 116 | 0 | 0 |
| Frente Amplio | 137 | 0 | 0 |
| ERC | 123 | 0 | 0 |
| Podemos | 74 | 0 | 0 |
| SALF | 0 | 0 | 0 |
| EH Bildu | 64 | 0 | 0 |
| Junts | 29 | 0 | 0 |
| Compromís | 16 | 0 | 0 |
| BNG | 4 | 0 | 0 |
| PNV | 20 | 0 | 0 |
| CC | 43 | 0 | 0 |
Neutrality
- Parties and their candidates always appear in the same order, from highest to lowest expected vote according to our polling average: PP, PSOE, Vox, Frente Amplio, ERC, Podemos, SALF, EH Bildu, Junts, Compromís, BNG, PNV, CC. Polls cannot be published from 24 to 29 November, and on those days the order is that of the vote in the 2023 general election.
- Same instructions and same model for everyone. The party's name is not part of what the model sees.
- Studies have found political biases in language models. That is why, before publishing, we swap the names of candidates and parties in equivalent excerpts and check that the classification does not change; and we compare, party by party, the model's classification with the human one.
- We do not show how many proposals each party has: that depends on how long its manifesto is, not on how much it cares about a topic.
Which interviews we analyse
Selection rule: [pending]. If a candidate has not taken part in an interview or debate, we say so in the calendar.
We analyse interviews and debates that have been publicly broadcast. For the analysis we download their subtitles or audio; we do not host or redistribute them. On this website, videos play in the broadcaster's official player, and we quote specific sentences, never the full transcript unless the rights holder allows it.
Election manifestos
The manifesto proposals are open data from 29N (29nov.es), under a CC BY 4.0 licence. The titles and explanations are written by 29N; the quote comes from each party's manifesto and links to its page.
Day of reflection
On 28 November, and on the 29th until the polling stations close, we publish nothing new and the AI assistant does not answer. The website remains open as an archive.
Who makes it
- Editor in charge: Rafa Conde.
- Development and funding: [pending].
- It accepts no money or sponsorship from parties, foundations or candidates.
- Errors and rectifications: info@29nresponde.es. We reply within 24 hours and record it in Corrections.
- Everything published can be downloaded from Open data. What data we process: Privacy.