Methodology

Pulse shows weekly counts of what this tracker collected. It is not a survey of any sector, party, or public: it measures what entered this tracker, not the world.

• Weekly counts. Each day we count items per week (weeks start Monday, UTC) by the kind of source, which we call a pillar: academia, industry, government, press, labor, civil society. Research and news are counted in the week they were published; bills are counted in the week of their most recent status change and include bills we track but have not published. The current week is partial. Weeks with fewer than five items of a kind are not shown. Weeks older than eight weeks are fixed and no longer recounted.
• Coverage rule. Coverage begins, for each kind of content, with the first calendar month that holds at least ten published items of that kind. Earlier weeks show no data, not zero, and are excluded from every average. The boundary is computed from the data on every refresh, never typed in by hand.
• Markers. Marked weeks reflect a change in this tracker, not a change in the field. Each marker is a dated change on our side, such as a change to the relevance check, a bulk publish, or a source outage, listed with its date under the chart it marks.
• Observed and reconstructed policy weeks. Bill status changes are observed as they happen only from the date this tracker first recorded a status change. Weeks before that date are reconstructed from the bill histories of bills our search found later, not observed as they happened, and bills that died early in 2025 are under-counted. The boundary is the earliest recorded observation, read from the status history on each refresh, and the reconstructed weeks are hatched on the chart.
• Impact-area tags. Impact-area tags are assigned when an item arrives, by Jev since 2026-09-30, and by a model (Claude Sonnet 4.6) before that. The relevance check was changed on 2026-07-22, which changed which items entered the tracker; shares before and after are not directly comparable.

Signals

Each published item and each tracked bill is read by a model (Claude Haiku 4.5) against a fixed taxonomy: six pillars for who speaks (academia, industry, government, press, labor, civil society, derived from the source, never from the model); an outlook (positive, mixed or neutral, concerned, not applicable); a regulatory stance (favors guardrails, descriptive or neutral, favors restraint, not applicable); up to three of fifteen topic priorities; and an evidence type (empirical, normative or theoretical, reporting, legislative text). The model scores the document’s own text only and quotes the span it relied on; a span that is not in the text rejects the score.

These scores come from a model, and they stay off the public site until two people have labeled the same items independently. A measure goes public only when the model agrees with the people closely enough (the low end of the confidence range for Cohen’s kappa, a standard agreement score, is at least 0.6), the two people agree with each other at least as well (kappa of 0.6 or more), and a test for bias shows a gap of no more than 15 points. Until then no stance, outlook or topic figure appears on Pulse.

Where a series is later extended backward from archives, the chart says which month those earlier figures were scored in, and that they were not observed as they happened.

The bellwether index

The long series on Pulse shows, from 2019, a monthly index across a fixed, named sample of archive sources, the bellwether index. It is not a census. The number of items behind each point is in the table under the chart. Earlier months are still being added.

Index definition: items are binned by calendar month on the source’s own date. Each source’s monthly count is divided by that source’s trailing-12-month mean, the mean of the twelve calendar months before the month over months that have any row; the index is written only when at least six of those twelve months have rows, so a source’s first indexed month comes after six months of history. 1.0 is that source’s normal pace. A pillar’s index is the equal-weighted mean of its sources’ indexes for the month, so no single source sets the scale; the table under the chart gives the contributing sources and the items behind each month, and the caption’s n is the item total. Annual-event sources (accepted-paper venues) are excluded from the monthly pillar index and shown separately as a count per calendar year per venue. Until the first pillar has an indexed month, Pulse shows weekly item counts for every pillar.

How sources are chosen: a source qualifies if (a) it publishes its own words or primary documents under a stable URL scheme with an archive reachable to 2019 or to the source’s founding, (b) excerpt-plus-link is permitted or the material is public record, (c) it is either a producer of the thing measured (papers, filings, rules, bills, statements) or a body whose statements move the producers. Press is excluded from the historical set by rule (no archive access). This is a judgment for now. A later statistical test of which sources move ahead of the others will replace it, so this is a list of candidates, not a validated set.

Where a source is described as leading, the claim rests on an out-of-sample test published on this page, re-run quarterly; sources that stop leading are removed.

Excluded by rule: general press archives; podcast audio and third-party transcripts; paywalled transcript vendors; social media posts (no stable archive, platform terms).

Party panel

Party is the party of the primary sponsor as recorded by LegiScan. Sponsor party is not the bill’s ideology. Bills with a Democratic and a Republican primary sponsor are counted once as bipartisan. Committee-sponsored bills, resolutions, Nebraska’s nonpartisan legislature, bills introduced before 2026, and states with fewer than 15 tracked bills are excluded. The seat-share tick is that party’s share of the bill’s chamber of origin (NCSL, 2026-01-27).

Tracked bills are the ones our LegiScan search surfaced and a model or person kept, not every AI bill filed. Counts include tracked bills not yet verified by a person. Search recall against the NCSL AI legislation list is 47% for the eleven states measured below.

The federal bills this tracker follows are a hand-curated list, not a search of Congress. Cosponsor parties are not yet collected.

Never a D vs R enactment rate. A raw comparison of enactment by sponsor party would show a gap, and the gap is a majority-status artifact, not a party effect: 26 of the 28 Republican-sponsored enactments in the 2026-09-18 party-facts snapshot sit in states with a Republican trifecta (Ballotpedia state trifecta list, read 2026-09-18). The rates themselves are not published. Enactment is therefore compared only as majority-sponsor against minority-sponsor within a chamber, and that comparison is not published yet.

We cannot yet say whether our relevance check turns away one party’s bills more often: only 2 of the 1,220 bills it turned away kept their LegiScan record. We will measure it once each party has 100 such bills on record.

Search recall per state against NCSL’s 2026 AI legislation list (retrieved 2026-09-19; a bill counts as 2026-introduced when its first recorded action is dated 2026; carryover bills are not scored). Computed by an AI agent against NCSL’s 2026 list; a person has not yet spot-checked it.

A state shown on Pulse that is not in this table reads “recall not measured”.

The ledger ranks AI events by public engagement on five free platforms. Each platform has one absolute bar, set by that platform’s own standard and read from the figures the platform itself reports, never estimated:

• Hacker News: a story about the event reached 500 points.
• OpenAlex: the work’s citations are above the 99th percentile of its publication year and above 50, and the work is at least 12 months old.
• YouTube: a video of or about the event reached 1,000,000 views.
• Wikipedia: a day in the 30 days after the event is both a spike (ten times the article’s trailing 90-day median) and in the daily top 1,000.
• GitHub: the repository gained 10,000 stars within 30 days.

C is the number of platforms crossed, 0 to 5; every event is tested on every platform where it can be looked up. Margin is how far past each crossed bar the event went, capped at ten times per platform, summed and scaled to 0 to 1; it only breaks ties inside a value of C. The score is C plus margin; ties after that go to the older event.

The weekly list needs one platform crossed and an event date inside the week. The monthly list needs two platforms crossed and, for the last twelve months, a place on a weekly list. The annual list ranks the year’s monthly survivors.

Before an event makes a list, two AI reviewers from different companies check it separately, without seeing its score or rank, against yes-or-no questions: is it relevant, is the engagement genuine, where did it start, is it a duplicate, and is the date right (the monthly list adds whether the event moved a government, civil society or labor source). Both must accept it. Each round includes two planted fakes, a repository with bought stars and a press rewrite of another candidate; if either reviewer accepts a fake, the round is thrown out and run again. We track how often the two reviewers agree.

Hacker News points, YouTube views and OpenAlex citations are current totals, not counts at the time, so recent years lag until the counts settle; Wikipedia views and GitHub stars are true time series. Lists are machine-ranked from these numbers with no editorial picks.

How to read these results

We do not test models ourselves. Every figure on the Assurance pages links to the tester who ran it.

• Each figure shows who tested it, when, how, and links to the original.
• Scores from different testers are never averaged or ranked together.
• A model a tester did not try is marked "not tested here", never shown as a low score.
• Consistency across repeated runs appears only where a tester measured it.
• Figures from AI vendors' own announcements are shown only in their own lane, never mixed with independent results.
• Where we do not have permission to show a tester's figures, we link to them instead.

Reliability and repeat runs

Reliability means how much a model's score moves when the tester runs the same items again. Where a tester did not publish that, the page says “Reliability not measured by this source.” beside the tester, so a single run is never read as a steady result. The fourth evidence ring needs repeat runs or an independent rerun. A tester that says it repeats its runs earns that ring even where we do not show the spread for each model; its method is described under the source.

Dates

A run date is the day the tester ran the model. Where a tester does not publish run dates, we show the date the results were posted and say so beside the tester. Where a tester does not say which kind of date it gives, we say that too. Results more than 12 months old are kept under “Older results”, apart from the current ones.

Quality and cost

Cost is shown only when the tester published it for the same run as the score. “Most accurate, and the cheapest within 1 point of the best” names the most accurate setup and the lowest-cost setup scoring no more than 1 point below it.

Sample jobs turn a share-correct score into errors per job. This is simple arithmetic on the test's score. Your documents will differ.

Findings and practices

Each finding holds only within the test it names. Cost appears only where the tester published it for the same run. On the pages, F marks a finding above and P a common practice: published advice from at least two independent sources, not measured here.

Input types

Each test is listed on its own. A file type counts for a test only when it is true of every item in that test, or of a published slice of 5 or more items. A smaller slice is shown as too few to rank.

Words used on these pages

Benchmark

A fixed set of test items with known correct answers, used to score different AI models the same way.

Field

One piece of information you ask the AI to pull out of a document, such as an invoice number, a date or one amount in a table row.

Line items

The repeated rows in a document, such as each line on an invoice or each check in a register.

Schema

The list of fields to pull out of a document and the form each answer should take, such as text, a number or a list of rows.

Answer key

The correct answers, checked by people, that each model's output is graded against.

Exact match

An answer counts only if it is the same as the correct one once small differences such as spaces, capital letters and punctuation are cleaned up.

Direct call

Sending a document straight to an AI model with the list of fields and taking the one answer it gives back, with no extra tools.

Coding agent

An AI model run inside a program that lets it use tools, such as searching a file or rereading pages, and check its own work before it answers.

F1 score

A score that balances two things: how much of what the model gave was right, and how much of what it should have found it actually found.

Hallucination

A value the AI made up that does not appear in the document.

Leaderboard

A published table that ranks AI models by their scores on a test.

Breakdown

The same test's results split into groups, such as by kind of file or by document length.

Reasoning effort

A setting on some AI models that lets them think longer before answering, usually at a higher cost.

Confidence interval

A range around a score that shows how much it could move by chance if the test were run again.