Seventeen topics, counted by week
Each row counts the Federal Register documents that match one written search, week by week, back to 2016. A topic is a subject, not a claim: the count says how much the government wrote about it, not what it argued, and it does not measure what people think.
The list was fixed on November 5, 2026, and no topic leaves it, whatever it shows. It is a judgment made on one day, not a sample, and it says nothing about topics left off it.
The latest reading
Through the week of October 26, 2026, retrieved November 5, 2026: 1 is growing, 7 holding, 2 fading and 7 too few to read. A reading compares the latest 12 weeks with the 12 before. Notes on each stored week.
Growing is up at least 25 percent, fading is down at least 25 percent and holding is in between. Too few to read means under 24 documents across both windows.
| 12 weeks earlier | 12 weeks before | Latest 12 weeks | Reading | |||||
|---|---|---|---|---|---|---|---|---|
| Critical mineralsSecurity and trade | 23 | −15% | +167% | 10 | Holding | |||
| Artificial intelligenceTechnology | 87 | −5% | +56% | 28 | Holding | |||
| Supply chainsSecurity and trade | 97 | −23% | +18% | 25 | Holding | |||
| Data centersTechnologyWithout SEC filings, which arrive in batches, 35 to 31: not a fade. | 57 | −38% | +13% | 21 | Fading | |||
| FentanylHealth | 28 | −7% | +9% | 6 | Holding | |||
| Unmanned aircraftTechnology | 22 | −12% | +4% | 10 | Holding | |||
| Critical infrastructureSecurity and trade | 25 | −38% | −4% | 13 | Fading | |||
| Digital assetsFinance | 32 | +3% | −6% | 5 | Holding | |||
| WildfireEnvironment and energyMuch of an autumn rise here is the season. This rise cleared 25 percent in seven of nine past years. | 48 | +66% | −9% | 15 | Growing | |||
| Foreign adversariesSecurity and trade | 12 | −14% | – | 8 | Holding | |||
| Quantum computingTechnology | 5 | – | – | 4 | Too few to read | |||
| Avian influenzaHealth | 5 | – | – | 3 | Too few to read | |||
| MeaslesHealth | 3 | – | – | 2 | Too few to read | |||
| PFASHealth | 3 | – | – | 3 | Too few to read | |||
| MisinformationInformation | 3 | – | – | 3 | Too few to read | |||
| Small modular reactorsEnvironment and energy | 2 | – | – | 2 | Too few to read | |||
| Bell to bellEducation | 1 | – | – | 1 | Too few to read | |||
| All Federal Register documentsFor scale | 5,877 | −5% | +3% | All | Not read |
The center column holds the latest 12 weeks. The dark bars to its left are the 12 before and the pale bars the 12 before those, and each row has its own scale. Each topic page shows the count back to 2016.
Documents is the count in the latest 12 weeks. The 12-week change is against the 12 before, and the 52-week change is the latest 52 against the 52 before. A dash means too few documents to compare. Publishers are agencies with a document in the latest 12 weeks. The 52-week figures still hold part of the 2025 funding lapse, as explained below.
Micro and macro, as last tested
Micro trends: 0
A micro trend has risen 3 weeks running with fewer than 10 publishers. No topic does so in the latest count. Three rises in a row turned up about as often as in shuffled counts. Replayed over the ten years to October 2026, this flag would have fired about 11 times, so it only says where to look closer.
Macro trends: none possible here
A macro trend needs publishers under different owners, in at least 3 regions or communities, and two kinds of publisher. Nearly every publisher in the Federal Register is part of one government, so this source cannot show one. On the two tests it can run, a document every week and 10 or more publishers, four topics pass: artificial intelligence, data centers, supply chains and wildfire. They are still not macro trends.
Backtest: what followed a reading
Method 1.5 says a reading describes the window just ended and says nothing about what comes next. Past readings agree. Used as a forecast, a reading was set against trivial rules in six comparisons fixed in advance. It was better in none, worse in two and not measurably different in four.
132 cases: nine topics on 16 dates, 12 weeks apart, from January 3, 2022 to June 16, 2025. None of these outcomes was used to fit anything. The question and scoring were written down before the first run. This measures past readings only. It predicts nothing, and nothing from it goes on the Record.
Naming the next reading
| Rule | Right | The reading against it |
|---|---|---|
| The reading repeats | 60 of 13245% | |
| Repeat the last window, which says holding | 70 of 13253% | No measured difference−7.6 (−16.3 to 0.0) |
| Same window last year | 57 of 13243% | No measured difference+2.3 (−6.1 to +10.3) |
Estimating the next count
| Rule | Error | The reading against it |
|---|---|---|
| The reading carried forward | 27.9% | |
| Repeat the last window | 24.0% | Worse+3.9 (+1.5 to +6.6) |
| Mean of the last two windows | 22.2% | Worse+5.8 (+1.4 to +10.5) |
| Pace of the last 52 weeks | 25.4% | No measured difference+2.5 (−3.4 to +8.7) |
| Same 12 weeks last year | 32.4% | No measured difference−4.5 (−12.2 to +3.6) |
Carried forward means up 25 percent after growing, level after holding and down 25 percent after fading, the smallest move each reading states. Error is all the documents missed as a share of all the documents counted, so lower is better, and larger topics weigh more. Under each result is the difference in percentage points, then its 95 percent interval, from resampling the 16 dates. A result counts as better or worse only if that interval leaves out zero. On a topic-by-topic count, the reading beat repeating the last window on one of nine topics.
The closest result
Against repeating the last window, the interval ends at exactly zero, so under the rule set in advance it stays no measured difference, by the smallest margin there is. The sample also flatters the reading. Its first date is one of twelve possible starting weeks, and it turned out to be the kindest. Across all twelve, the reading was right 19 to 45 percent of the time, against 40 to 58 for the repeat, and it trailed the repeat at every one of them, on both questions. Against the repeat, the interval left out zero at eleven of the twelve when naming the reading, and at twelve when estimating the count. The reading was better at none. Over all 1,535 weekly cases, which overlap heavily, it was right 36 percent of the time against 51.
What followed each reading
| Reading | Cases | Growing | Holding | Fading | Too few |
|---|---|---|---|---|---|
| Growing | 33 | 7 | 14 | 12 | 0 |
| Holding | 74 | 13 | 50 | 10 | 1 |
| Fading | 25 | 14 | 6 | 3 | 2 |
| Any | 132 | 34 | 70 | 25 | 3 |
Holding was the one reading that tended to repeat. Fading was followed by growing 14 times in 25. Much of that is arithmetic: a high window is the base of the next comparison, so a rise is hard to repeat, and a rise from a low window comes easily. Since 2016, fading was followed by growing 54 percent of the time, and 53 percent with each topic’s weeks shuffled.
- The 13 percent
- Trends says a growing reading was followed by another 13 percent of the time. That stands: 143 of 1,133, or 12.6 percent, over every week and topic through the week of October 26, 2026. Those 1,133 cases are 235 separate runs of weeks in eleven topics, because one rise reads growing for weeks on end. Growing was 30.1 percent of all readable cases, so a reading right after a growing one was growing far less often than readings are in general. With each topic’s weeks shuffled, the rate is 8.6 percent (5.9 to 11.4), so the real rate is a little higher than with the order removed. In the held-out weeks it is 50 of 414, or 12.1 percent. The sample in the tables gives 7 of 33, or 21.2 percent, and the twelve starting weeks run from 6.3 to 21.2, which is how far so few cases can move.
- A rise that stayed
- Failing to grow again is not the same as falling back. After a growing reading, the next 12 weeks were higher still 13 times in 33, and they stayed at least 25 percent above the window before the rise 19 times in 33. Over all weekly cases those are 27 and 54 percent. Neither has a baseline, so neither is a test. Since 2016, growing was followed by fading 32 percent of the time, and 42 percent with the weeks shuffled, so rises held better than noise would.
- Sample
- History was split by time at January 3, 2022, not at random. Cases are 12 weeks apart because weekly ones overlap, but neighbors still share a window, and many topics move on the same day, so the intervals treat each date as one unit. 16 dates is few. Of 3,145 held-out topic-weeks, 1,610 were too few to read and made no call, and eight of the seventeen topics have no case in the sample. Nothing here speaks for them.
- Not blind
- “Held out” means held out from what this test fits. It does not mean unseen. The topics were chosen in 2026, knowing what had come to matter, and the window and floor were set after the first counts were seen.
- Funding lapses
- Any case that touches a week of the 2025 lapse is left out, so the sample stops before it. Put back, the sample grows to 178 cases. The reading still trails repeating the last window on both questions (46 percent right against 56), and both intervals leave out zero. The lapse from December 2018 to January 2019 also cut the count, and it is not excluded. It falls before the split, so it touches the 13 percent and the development period, not the sample in the tables.
- Correction to the plan
- As first written, the plan gave weeks for the last-year rules that sat 12 weeks later in the calendar than the window they stood for. That was found after the first run and fixed. The first version gave 57 right and 32.4 percent error. No verdict changed, and both versions are in the results. The plan lists every change made to it.
- Download
- Pinned to the counts through the week of October 26, 2026, so new weeks do not move it: every result, every weekly case and the plan.
How this is counted
- Source
- The Federal Register, the daily journal of the United States government, through its public search. Each topic page shows the exact search and a link to run it for the latest window. A document counts once, however often the words appear in it.
- Other sources
- The same topics are also counted in company filings and state bills, as separate series. Those counts are never added to these.
- Reading
- The latest 12 full weeks against the 12 before them, at 25 percent either way, with a floor of 24 documents. The window was set before the first run of the full list and the floor after it, and these numbers are proposed. A shorter window can give a different reading. The rules are in Method 1.5, where these readings are called Grow, Hold and Fade.
- What a reading is not
- It describes the window just ended, not what comes next. Since 2016, a growing reading was followed by another one 12 weeks later 13 percent of the time, and the backtest gives the fuller measure. It does not adjust for the time of year. Once or twice a year the government’s regulatory agenda mentions many of these topics at once. On August 14, 2026, it mentioned 11 of them, and a reading moves when such a day enters or leaves a window.
- The 52-week change
- The latest 52 weeks against the 52 before, with no percentage under 104 documents. Each side holds every season. This comparison and its floor were set after the first counts were seen, and these numbers are proposed. A 12-week window is not compared with the same 12 weeks a year earlier, because the funding lapse of October 1 to November 12, 2025 cut the Federal Register to about a tenth of its usual weekly count, and a catch-up followed. This comparison is not free of the lapse either: 4 of its 8 low and catch-up weeks are in the latest 52 and 4 in the 52 before. That moves these percentages, most of all the last row, until those weeks leave the comparison. The share on each topic page allows for it.
- For scale
- The last row counts everything published. A topic can rise only because the whole Federal Register grew, so each topic page also gives its share of all documents.
- Wording
- A topic is one wording. Other words for the same thing are not counted, so a shift in official vocabulary can look like a fall. Topic pages name the main gaps.
- Publishers
- Top-level agencies with at least one document in the latest 12 weeks. A bureau or office is folded into its department, following the Federal Register’s own agency list.
- Not checked
- Each document is counted as the source returns it. No copy check is run and no item has been audited.
- Updates
- The stored counts come from one run by hand on November 5, 2026. A job is set to add each finished week on Mondays, checking the latest 12 weeks of each topic against a separate direct search before it stores them. There is one stored run so far, so no past week has been compared yet.
- Searches corrected
- November 5, 2026. The source treats a plural as a different phrase. “Wildfire” alone returned 28 documents from August 10 to November 1, 2026, and both forms returned 50. The searches for data centers, critical minerals, supply chains, foreign adversaries, wildfire, small modular reactors and digital assets now list both forms.
- November 5, 2026. “Data center” also caught the National Flight Data Center, an FAA office named mostly in routine flight-procedure rules: 12 of the 69 documents with either form from August 10 to November 1, 2026, which leaves 57. That name is now left out of the search.
- November 5, 2026. “Wildfire” also caught a company called Wildfire Energy in merger notices: 2 documents from August 10 to November 1, 2026. That name is now left out of the search. The same exclusion also drops two 2016 Interior notices where wildfire and energy sit side by side in a list.
- Download
- Every weekly count, for anyone to check: CSV or JSON. The backtest has its own downloads.