How we decide
Rubric version 1.0-mock · updated Oct 31, 2026
Sample data. The rubric version, model names and test-set figures here are placeholders.
The rubric
What the scorer can’t see
The scorer receives only the bill’s title, official summary and section text. Sponsor names, party, state, committee and member names inside the text are removed before scoring and added back only for display.
Which bills are scored is set by a fixed rule: every bill that has been reported, passed a chamber or been scheduled for a vote. Nobody picks bills by sponsor.
The exact prompt
Published verbatim for rubric v1.0-mock. Anyone can run it on the same text and compare.
Placeholder. The exact prompt is published here verbatim by scoring/ for each rubric version.
Models
| Role | Model and version |
|---|---|
| Scorer | to be published by scoring/ |
| Consistency check (second pass) | to be published by scoring/ |
| Second-model sample check | to be published by scoring/ |
Every flag records the rubric version and model it was made with.
Human review
A person approves every flag before it is published. Reviewers can approve or reject what the scorer produced, with a logged reason; they can’t add flags of their own. Flags on bills whose sponsor list is from a single party need two reviewers.
Test set and accuracy
200 labelled sections. Agreement with a person’s call: not yet published.
We also check that rewording a section with left- or right-leaning language doesn’t change the call, and re-score a sample with a second model from a different provider. Monthly results by sponsor party are on the party audit page.
Changes to the rubric
When the rubric changes, every bill in the current Congress is re-scored with the new version.
- v1.0-mock (Oct 31, 2026): Placeholder for the first published rubric.
Think a call is wrong? Challenge a flag. Accepted challenges are listed on the corrections page. The full neutrality standard is published too.