How Apogee Insight Pages Are Made
Data sources, how lobbying data is matched and its limits, the automated checks that run before an insight page is published, refresh cadence, and how to send a correction.
How Apogee Insight Pages Are Made
Insight pages are short, sourced explainers on a live policy topic, built from primary records. This page states what runs before an insight page goes live, where the data comes from, and where the numbers stop being reliable.
Accountable editor: Shaun Brown. Corrections: reach@apog.ai.
Insight pages are produced by Apogee research agents from primary records. Automated checks run before publication, and an independent AI verifier checks the claims where the page says it ran (see Independent verification below). Not reviewed by a human before publication.
Who is accountable
Each insight page names an accountable editor. That credit states responsibility for the page and for answering corrections. It does not mean the person read the page before it went live: no per-page human review takes place today.
Data sources
- Federal bills, actions and text: Congress.gov (read through its API when a page is generated) and the Government Publishing Office.
- Federal rules and dockets: the Federal Register.
- Lobbying: quarterly filings under the Lobbying Disclosure Act, from the Senate lobbying disclosure database.
- State bills and status: state legislature records, read through the Open States API where available and otherwise from a synced copy in the Apogee Policy Graph.
Every page lists the sources it was built from. Links in the text point only to those sources.
Reading bill tables
"Last action" is the newest entry in a bill's record. It is not necessarily a vote or a final outcome. Each row links to the official page, which is the current record.
How lobbying data is matched, and its limits
A lobbying filing is tied to a topic when the text in its specific-issue field names the topic, for example "data center". It is tied to a bill when that text names the bill. Limits that apply to every number built this way:
- Filings are self-reported by registrants, and dollar figures are reported per registrant per quarter, not per issue.
- A filing that mentions a topic may cover many other issues. The mention can be incidental.
- Text matching misses filings that use different words for the same subject.
- Filings are published after each quarter closes, so the newest activity is not yet counted.
Limits of state data
State records can lag the official legislature site. A bill status on a page is the status as of the page's "Data as of" date, which may differ from today's status. When a status cannot be read, the table shows the bill's last recorded action instead. Whether a data center bill is a pause is decided by keyword rules on its official title, not by reading the bill text. The official record always controls.
Automated checks that run before publication
Pages are written by AI research agents that read the sources above through tools. Before a page is published, automated checks run with no person in the loop, and a page that fails is not published:
- The page must parse into the required structure, with sections, at-a-glance figures and at least one cited source.
- Links must point to the page's own sources. A link to anything else is reduced to plain text.
- Bill links must match the bill named in the link text.
- Claims such as "enacted" or "signed" must be backed by the status field of the record the agents read.
- Known wrong-phrase patterns, such as leaked working notes, are removed.
These checks catch structural and mechanical errors. They are not a substitute for reading the official record, and they can miss a fact that is wrong or out of date.
Independent verification
A second AI agent, from a different vendor than the writing agents, checks the claims on a page against the primary records before publication. The disclosure at the top of each page states whether that independent check ran for that page. It is an AI check, not a human review.
Refresh cadence
The target is a weekly refresh. The dates at the top of each page show when it was first published, when it was last regenerated, and the date of the data it reflects. A page that is not refreshed within its window is withdrawn from search.
AI disclosure
These pages are AI generated. Text, figures and structure are produced by AI agents from primary records, then passed through the automated checks above.
Corrections
To report an error, write to reach@apog.ai with the page address and the statement in question. Corrections are made by regenerating the page from corrected data or by editing the checks that missed the error.