The Daily Trickle · Dashboard

How we measure

The Daily Trickle measures the gap between what matters and what gets covered. Every day it reads what governments and scientists publish, filters it down to what is newsworthy, scores each document's impact, measures how much the news covered it, and puts the high-impact, low-coverage items first. This page shows every step, how good each measurement is, and where it is still weak.

Numbers as of November 8, 2026: 1759 leads published in 49 batches (530 rules, 1134 research papers, 67 institution announcements and 28 public-health reports).

1. Sources

Only primary documents, from official APIs and feeds, within their rate limits and robots.txt. No news outlets are read or summarized.

• Rules: the Federal Register (every final and proposed rule); the state registers of California, Oregon, Texas and Washington; and from the United Kingdom, statutory instruments from legislation.gov.uk (drafts laid before Parliament, and instruments as made).
• Research: journals and preprint servers (Nature, Science, Cell, PNAS, PLOS, arXiv, bioRxiv, medRxiv, ChemRxiv, NBER), through their feeds and Crossref.
• Reports and announcements: the CDC's MMWR, and the newsrooms of a short list of widely trusted institutions.

2. Filtering: what reaches the page

1. Duplicates are dropped: the same document id or a near-identical text at ingest, and the same paper (DOI) arriving from a second source.
2. Code rules drop routine documents before any model sees them:
   • Corrections: a correction to an earlier document, or a UK instrument re-made to fix one
   • Technical amendments: fixes to wording, addresses or cross-references
   • Comment-period changes: extensions or reopenings of a comment period
   • Hearing notices: notices that only schedule, move or reformat a hearing on an earlier proposal
   • Airworthiness directives: orders about one aircraft model
   • Single-entity actions: one state plan, one product, one station, one airport's airspace, one bridge or zone; in the UK, one flying restriction, one council's ward boundaries or one stretch of the coast path
   • Not a rule: documents that are not final or proposed rules
3. A model gate (Jev, a decision model) asks of each remaining rule: would a non-specialist in the affected group find this notable? A rule is kept unless the model puts that chance at 30% or less. Re-scored weekly on 208 hand-labelled rules (latest 2026-11-02), the gate's precision is 0.74 and recall 0.81: it keeps most of what a person would keep, and lets some routine items through rather than lose real ones. Research has its own question, not yet calibrated against hand labels.
4. Ranking and slots. Each daily edition writes up the top-ranked documents, up to 40 a part, half for rules and half for research and the rest (their scores come from different questions, so they are ranked apart). The others wait for a later part.
5. A fact check (code, not a model) holds back any summary whose dates, numbers or names are not in the source, that calls a proposal final or a preprint peer-reviewed, or that names a person the document does not.

The run dashboard shows this funnel live, with the counts at every step.

3. Impact and the Trickle score

Impact is a model's judgment of how many people a document affects and how much, from None to Sweeping. The Trickle score (0 to 100) combines impact with how little the news has covered the document: the geometric mean of impact (0 to 1) and 1 / (1 + log10(1 + stories)), so 0 stories gives 1, 9 give 0.5 and 99 give a third. A lead must do well on both to rise. Interest (how surprising it is) is shown but kept out of the score until it has been checked against hand labels. A lead without a coverage count is scored on impact alone and drawn paler.

4. Coverage: how much the news said

We count matching stories with Media Cloud, an open-source platform for media research. We show only the number, never the stories.

• Where: federal actions in US national news; state actions (California, Oregon, Texas and Washington) in that state's state and local news; research, announcements and reports in US national news (experimental: not yet checked against an independent search). UK statutory instruments are not counted yet: their own country's news is where they would be covered, and it is not set up. Counted so far: 1358 in US national news, 129 in Oregon state and local news, 111 in Washington state and local news, 56 in Texas state and local news, 45 in California state and local news.
• The search is built from the document's own words, checked by code: its most distinctive phrase (a program, a species, a statute) and one or two context terms such as the agency, with agency names widened to how the news writes them ("HHS", "Department of Health and Human Services"). A formal title is also searched in its plain form ("Crypto Custody Rules" or "Crypto Custody"), and when the search names an agency, the agency must appear: a loose topic word cannot stand in for it.
• Companion rules: agencies often publish a package of rules on one day, and the news covers it under the main rule's name. A silent rule from the same agency on the same day that shares the main rule's key word shares its count, and says so.
• The window runs from 7 days before publication, when agencies often announce, to 14 days after. Counts stay open ("so far"): a lead is counted when it is published, then again on days 0, 1, 3, 7 and 16 after the document's date (research on days 1, 7 and 16); after day 16 they are final and never requested again.
• Zeros are tested, not trusted. When the search finds nothing, we check whether its key phrase has appeared in the news at all in the year before. If it has, the zero is real silence. If it has not, we try a plain-words search; if that also finds nothing the lead is silent, and otherwise it is marked "search unreliable" and left out of the score.
• Every lead's "How this was measured" shows its search, window and result.

Of 479 counted rules, 427 have a count we trust (89%), and 341 of those (80%) have had no story at all so far. Research, announcements and reports (experimental): 1055 of 1220 have a count we trust, and 928 of those have had no story in US national news.

5. How good is the coverage count?

On 2026-10-09 we checked every counted federal lead (58) by hand: an independent web search for US national news coverage of that action. We keep counts and verdicts only, never outlets or links.

Counting method 2: recall 60% (we found coverage for 12 of the 20 covered actions) and precision 71% (of our positive counts that the check could settle, that share was real coverage). The first check, of method 1, found recall 58% and precision 58%; method 2 opened the window a week before publication and searched agency names as the news writes them.

What the check found:

• The earlier window caught 7 actions covered at the agency's announcement, which method 1 missed entirely.
• Most remaining misses come from search terms: formal rule titles the news never uses.
• Companion documents published with a main rule are covered under the main rule's name.
• Counts of 1 or 2 are still the main source of false positives.
• State searches with generic terms match unrelated local news; the state sample was too thin to measure recall.

State actions (20 checked against that state's own local news): 1 had any local coverage, so state recall cannot be measured yet; precision was 25%, because generic searches (a program name and the state) match unrelated local stories.

Recall is still well short of the 80% we aim for. Counting method 3, live since 2026-10-10, searches formal titles in their plain form, lets a companion document share its main rule's count, and requires the agency to appear when the search names it. It has not been checked yet; the next recall check measures it.

Quality at a glance

Every step's accuracy against its target. Gate and score metrics are re-scored weekly on hand-labelled sets; coverage recall and precision come from a hand-checked web search, monthly and after any change to how coverage is counted. The run dashboard shows the trend.

6. What is still weak

• Coverage of research is new and experimental, only two states are counted, and UK leads are not counted.
• Coverage recall is 60% on the latest check, short of the 80% target, and a count of 1 or 2 is often a different story on the same subject.
• Impact, the research gate and interest are model judgments still being calibrated against hand labels.
• The site is new: it went live on October 7, 2026.

Built by one developer with a code pipeline and AI models; every summary links to its source document. Not legal or medical advice.