Did the left say it, or the right?
A mobile game built on real quotes from the Spanish parliament: you guess which side said each one, and the card shows who it was, with a link to the official record. The interesting part is behind it: no quote is picked by a person, and I measured the model that filters them against my own blind review. It's live at escanometro.com, in Spanish.
- Role
- Idea and product and method decisions; the code and much of the analysis with an AI agent I directed
- Tools
- Python, qwen3-32b run locally, Claude Sonnet and DeepSeek as judges, Cloudflare Pages, D1
- Result
- Live since 5 October 2026, with the method published. I never got round to promoting it, so there isn't yet a sample to report figures
The problem
We think the two political blocs are further apart than they are. Social psychology calls it false polarisation, and one of its causes is published: we generalise about "what the left says" or "what the right says" from evidence that doesn't support it (Novoa et al., 2023, PNAS). I wanted to test it against what is actually said in the Spanish parliament: how many quotes someone who follows politics would place wrongly, and above all how many they'd place wrongly while feeling sure.
A tool like this only works if nobody can accuse it of picking quotes to suit one side. On a political subject, the defence can't be the good faith of whoever built it. It has to be checkable.
The solution
A ten-quote game. A quote from the parliamentary record appears and you bet with four buttons: "sure left", "I think left", "I think right" or "sure right". When you answer, the card shows who said it, their group, the date and a link to the official PDF of the session. At first confidence was a separate question; in my first test I answered "sure" to all twenty quotes and the extra question slowed the game down, so it moved into the buttons.
No quote is picked by a person. The catalogue comes from 76,421 speeches in the current legislature (824 sessions), through filters written in code:
- length, topic by keyword list and Spanish only, because a translation stops being a quote;
- out goes anything that gives the speaker away without an argument (party names, "our group", "minister") and anything quoting someone else, because an MP quoting a rival would produce a false attribution with correct data;
- four groups (PP and Vox, PSOE and Sumar) and no others, because if the right answer is debatable the tool measures nothing.
Playing it, I found quotes that gave themselves away: names of politicians, regional party acronyms, Catalan political terms. None was removed by hand. Each pattern became a general rule with its own test, and one filter that hit one bloc harder than the other was dropped.
A language model, measured before use. An open model, run locally, decides whether each quote is really about the topic and whether it makes a claim you could argue for or against. It never sees the party or the speaker. To choose it, I blind-reviewed 50 quotes myself, seeing only the topic and the text, and compared three models with my review. The first round showed that my definition of "stance" was ambiguous: I rewrote it and repeated that part. The model I chose agrees with me on 90% of quotes about whether they're on topic and 86% about whether they take a stance. It's poor at spotting filler; I accepted that as a risk and the methodology says so.
Every quote checked against the official PDF. 238 of the 240 by two methods, text and speaker, and the other two by one.
The method, published before any result was analysed. The methodology is on the site. The rule for retiring a quote as too easy (15 answers and 85% correct) was set on 4 October, before the first answer.
Privacy by design. No cookies, nothing stored on the device and no question about the player's own views, only their attribution. That way the answers aren't data about anyone's politics.
Result
The game has been live since 5 October 2026, with 240 quotes (40 per topic and bloc, on pensions, health and housing) and the methodology published.
A test game, not anyone's result.
The goal was for it to spread, and I never got round to promoting it. The first forum on the plan required new accounts to build up some activity first, which I didn't do, and I didn't push it through other channels. With the games played so far, I can't give the share of misattributed quotes without the figure claiming more than it knows, so I don't.
How it's built, for anyone who wants the technical detail
- Data: I start from the open parse of the parliamentary record by the escano-abierto project (MIT licence), from 17 August 2023 to 23 September 2026. The pipeline is in Python, standard library only.
- Selection: 3,724 candidate quotes after the filters, 3,227 after the model, 1,478 approved by both quality judges and 240 in the catalogue, by random sampling with a fixed seed and the same number of quotes per topic and bloc.
- Quality: two independent models (Claude Sonnet and DeepSeek) judge each quote, and the benchmark for how many faulty ones slip through is a blind reading by a third model.
- Web: Cloudflare Pages and a D1 database for the answers, with no server of its own.
- Code: MIT licence, at acamposla/escanometro, with 60 tests and the method explained step by step in the README. Code and docs are in Spanish.