How USFedPolicy works

Where the documents come from

The Federal Register's public API lists each new rule, proposed rule, notice, and presidential document. When a document has no summary, the start of its text is taken from GovInfo, the Government Publishing Office's site. This site checks for new documents every hour.

How a document is scored

Each document is sent to Jev, a model from TypeSafe AI that returns structured answers with probabilities instead of writing text. Jev answers four questions about every document, in a single request.

Is it routine?

Is the action described in `document` routine or administrative, such as a meeting notice, an information collection renewal, a technical correction, an approval limited to one place or one company, an ordinary permit or grant, or a ceremonial observance?

If Jev judges a document at least 50% likely to be routine, it is not placed on the scales. Otherwise it gets a reading on each of the three scales below.

Democratic to Republican

Which U.S. political party's national platform is the action described in `document` more consistent with? Judge what the action does, not who issued it.

1. 1 Strongly consistent with the Democratic Party platform and opposed by the Republican Party platform
2. 2 Somewhat more consistent with the Democratic Party platform
3. 3 Consistent with both party platforms, or consistent with neither
4. 4 Somewhat more consistent with the Republican Party platform
5. 5 Strongly consistent with the Republican Party platform and opposed by the Democratic Party platform

Economic left to right

Does the action described in `document` move U.S. economic policy toward the economic left or the economic right? Judge what the action does, not who issued it.

1. 1 Strongly economic left: greatly expands government regulation of markets, taxes on business or wealth, public spending, public ownership, or worker protections
2. 2 Somewhat economic left: modestly expands regulation of markets, taxation, public spending, or worker protections
3. 3 Economically centrist or mixed: no clear change, or expands in some ways and reduces in others
4. 4 Somewhat economic right: modestly deregulates markets, reduces taxes or public spending, privatizes, or reduces worker protections
5. 5 Strongly economic right: greatly deregulates markets, cuts taxes or public spending, privatizes government functions, or removes worker protections

Libertarian to authoritarian

Does the action described in `document` increase government control over individuals, or increase individual freedom from government? Judge what the action does, not who issued it.

1. 1 Strongly libertarian: greatly expands individual rights or freedoms, or greatly limits government power, surveillance, or enforcement over individuals
2. 2 Somewhat libertarian: modestly expands individual rights or freedoms, or modestly limits government power over individuals
3. 3 Neutral or mixed: no clear change in government control over individuals
4. 4 Somewhat authoritarian: modestly increases government control, enforcement, surveillance, or restrictions over individuals
5. 5 Strongly authoritarian: greatly increases government control, enforcement, surveillance, or restrictions over individuals

Jev returns a probability for each level, and the reading is the probability-weighted position between the levels, shown as a slider. Averages and counts on this site are calculated here, in code, not by Jev.

What the scores do and do not mean

• A reading describes what an action does, as read from its summary or opening text, not who issued it or why.
• Democratic and Republican refer to how well an action matches each party's national platform. They say nothing about whether an action is good or bad.
• Jev reports how confident it is in each reading. High confidence means the model is sure of its answer, not that the answer is correct.
• The model reads the summary, or at most the first 5,000 characters of the text, so a long document may be judged on its opening alone.
• Scores can be wrong. Use them to find documents worth reading, and read the source before drawing conclusions.

Weekly recaps

Every Monday the site publishes a recap of the previous Monday to Sunday week. Recaps are assembled from counts, averages, and rankings of the stored scores, so every sentence is a statement about the data. No text is generated by an AI model.

Privacy

There are no accounts, comments, cookies, or analytics on this site.