A political detection codebook first filters out non-political content. Three additional codebooks define the target constructs for the research projects: gender analysis (gender, valence, rationality, science position), technophile neo-reactionaries (technology, libertarianism, fictional metaphors, equality, ecology) and SIED (nationalism, immigration, democracy, progress, authority, tradition, equality, ecology) following the framework developed in Boursier & Lemor (2025). The LLM receives each codebook as a system prompt and annotates batches of 500 sentences, producing labels and justifications.

For each classification task, a CamemBERT model is fine-tuned on the LLM-annotated data. Models are evaluated on a held-out validation set. If the F1 score is insufficient, the pipeline enters a reinforced loop: the most uncertain predictions are sent back to the LLM for re-annotation with stricter prompts, the training set is augmented, and the model is retrained. This loop typically converges within two to three rounds.

Validated classifiers run inference on the entire corpus. Each of the 20 million sentences receives a prediction and a confidence score for every applicable model. All results are stored in PostgreSQL, enabling queries such as "show all political sentences annotated immigration_security" or "compare gender valence across political orientations".