AI Tribune

Public methodology

How this register is built

The unit of record

This register does not track bills. It tracks asks. An ask is a specific requested outcome, like “preempt state frontier-safety statutes.” Each ask answers a stable question, and each ask is evidenced by claims: one actor asserting one ask in one document. Bills, rulemakings and consultations are vehicles an ask rides. The moratorium was stripped from H.R. 1 by a 99 to 1 vote in July 2025, and the ask it carried was back inside an executive order within five months. Bill-status called that dead. Ask-status recorded what actually happened.

The verbatim rule

A claim without a verbatim quoted span and a page reference does not exist in this register. The extraction pipeline enforces this in code: a quote that is not an exact substring of the source document is rejected at write time. This is the one constraint that makes an automated register checkable by a stranger. You can click any claim and read the actor's own words in the actor's own document.

The daily pipeline

  1. Harvest. Poll every source adapter. Hash and archive each new document.
  2. Triage. A cheap classifier asks two things: is this about AI governance, and does someone ask government to do something? Most documents die here. The kill rate is logged per source, because a sudden change in it means a source broke.
  3. Resolve the actor. Exact match against the canonical actor table, aliases included. Unresolved names go to a queue. Nothing is fuzzy-matched into the record.
  4. Extract claims. What is this actor asking for, in what words, at what page, with what framing. The verbatim rule applies.
  5. Map to an ask. Above a confidence threshold, the claim attaches to an existing ask. Below it, the pipeline proposes a new ask and waits. It never creates one on its own.
  6. Detect. Across the whole corpus: the same actor holding incompatible positions in different forums, safety language wrapped around delay requests, near-identical statutory text across state bills, submitted language surfacing in enacted text, and silences from actors who normally file.
  7. Publish or hold. High-confidence claims go live. Everything flagged waits for a human. Detection findings never publish themselves.

The human loop

About half an hour a week: approving or rejecting proposed asks, clearing the entity queue, signing off findings before they publish, and revising the question layer quarterly. The ontology is the one part that cannot be automated. It is also small. That is the trade.

What this register will not do

No scores and no hypocrisy rankings. The moment you rank actors, every record you publish reads as an accusation, and the record stops being usable by both sides. No motive language either. The register states what was filed, where, and by whom. Why is your question to answer, with both documents in front of you.

The bias we can't remove

Formal filings oversample the side with compliance budgets. A trade association files forty comments a year. A worker affected by the same rule files none. We cannot fix that imbalance, so we state it: every major instrument page carries a coverage note saying what was indexed, from whom, and which channels this register cannot see.

Accountability surfaces