State PAP starter kit
Interested in applying the policy agendas codebook to a domain the project does not cover yet? This page is how you start one. The policy agenda is not only national: below is what it takes to stand up a state-level project on the same topic system as the national series — and which parts already exist, so you do not build them twice.
Six steps
Use the CAP master codebook unchanged, plus a code for material with no policy content. Resisting the urge to invent state-specific topics is what keeps a state comparable with the national series and with every other country project. A state's own questions live in the subtopics, not in new major topics.
Four source families carry a state agenda: the legislature (bills, and the laws that pass), the governor (State of the State addresses, executive orders), the budget, and the state press. Press is the cheapest to start with and the easiest to keep current; bills and laws are the most comparable across states.
Collect, deduplicate on content rather than URL, classify, then check. The classifier baseline for CAP English is the pair of fine-tuned XLM-RoBERTa models (English v5 and pooled v4); a general-purpose language model votes beside them only where they are close, and never codes alone. A state coded by a language model alone is drift rather than data. Every row keeps the model that coded it, so “what produced this code?” always has an answer.
Three checks are worth enforcing from the first week, because each catches a failure that otherwise looks like data: coding coverage (an uncoded backlog silently biases every share you compute), classifier provenance (what share of coded rows came from the baseline model), and jurisdiction — a state feed carrying another jurisdiction's press means rows were stamped upstream with the wrong project. That last one is not hypothetical: it is how roughly 1.1 million Australian articles were found sitting in two US state feeds.
Before a state's numbers carry an argument, double-code a sample by hand and report agreement. The classifier being fast is not a reason to trust it; the project treats inter-coder reliability as the gate on a gold standard, and a state project should hold itself to the same bar.
From this side: the codebook and its training material, the topic system and its labels, the classifier’s public model card, and a description of the pipeline to follow. From yours: source access, the judgement calls a state's own politics require, and the hand-coding that validates the result.
Thinking about starting one? Write to uspolicyagendas@syr.edu — the project's address, and the right first stop. Say which state and which sources you can reach; that is the useful conversation, and it comes before any coding. Other ways to reach the project are on the contact page.