A missing endpoint is the easy failure. The harder one returns HTTP 200 with fewer columns than it had last year. A new federal data tracker now lists 375 records and labels 38 data products as terminated. Many of the remaining records document narrower changes, such as removed survey questions, cancelled follow-ups or public interfaces that no longer expose the same material. For software built on government data, that makes policy change look uncomfortably like schema drift.
The public snapshots do not present the same total. The tracker said 28 datasets were deleted and 338 collections had been modified when The Guardian reported on it on August 18. The tracker page, which carries an August 17 update date, presents 375 entries and 38 terminations. The mismatch is a reason to treat the database as a versioned inventory, rather than a final census of everything removed since President Donald Trump returned to office.
That distinction matters. A terminated survey, a historical file moved behind a different access route and a live survey stripped of one variable produce different failures. One ends a time series. Another breaks a downloader. The third may leave a pipeline running while changing what its output can support. The Federation of American Scientists' methodology says the tracker counts at the product level, ties terminations to primary evidence and excludes temporary takedowns that have been restored.
The visible total hides several failure modes
The tracker defines termination narrowly enough to avoid counting every vanished web page as a lost dataset. Its criteria cover the documented end of a primary federal collection, the discontinuation of a derived product that cannot readily be recreated outside government, and the removal of substantive elements from a continuing product. Every counted termination must be tied to an agency notice, regulatory filing or other primary evidence, according to the published methodology.
That approach also explains why the 375 entries should not be described as 375 deleted datasets. The tracker includes caveats beside individual records. The National Household Education Surveys Program's 2023 files, for example, were advertised but never released; earlier files remain available. The National Teacher and Principal Survey is roughly 30 months past collection, but its public-use releases have historically taken 24 to 36 months. Those are warning signs and missing releases, not proof that every underlying record was destroyed, as the tracker itself notes.
At the other end is an unambiguous stop. The US Department of Agriculture announced in September 2025 that it was ending future Household Food Security Reports. The USDA release called the work redundant and politicized, and said the department would use other, more timely datasets where needed. The tracker describes the associated Current Population Survey supplement as the basis for the department's annual estimate of how many households struggled to afford enough food. No compatibility layer can replace a future observation that was never collected.
A third case leaves the collection alive while changing public access. On August 14, the Consumer Financial Protection Bureau stopped publishing complaint narratives and its visualizations. The bureau said the narratives were unverified and could mislead consumers. It will keep collecting complaints and place earlier narratives in its Freedom of Information Act reading room, but the tracker's caveat says those records will no longer be available in bulk. A researcher may still retrieve a document. A service that depended on a machine-readable corpus has lost the product it used.
A successful request can still return a broken series
Developers usually design for transport failures: a 404, a timeout, a changed hostname or an expired credential. Several entries in the tracker point to semantic failures instead. The data arrive, yet the meaning or population coverage differs from the earlier series. Unless a pipeline checks fields and distributions, it may publish a clean chart that crosses an undocumented boundary.
Gender identity and sexual orientation fields are the clearest recurring example. The White House's January 2025 executive order directed agencies to use "sex" rather than "gender" in applicable documents, remove messages that promote what the order calls gender ideology, and stop requesting gender identity on forms that require sex. The tracker records resulting field removals across health, justice, housing and other collections. The policy source is explicit; the technical consequence is that a previously observed subgroup can disappear from new rows.
The affected products are not obscure administrative forms alone. The tracker lists the CDC's National HIV Surveillance System, the National Violent Death Reporting System and the Medical Monitoring Project among continuing systems that lost gender identity items. In its August 18 analysis, The Guardian found health records made up about 40 percent of the deleted or altered collections then tracked. A model trained on older releases and refreshed with newer ones may therefore encounter a category that appears to go to zero because collection stopped, rather than because the population changed.
Historical comparability can fail without any column disappearing. The tracker says the Bureau of Prisons stopped reporting gender identity and that the relevant live report was not archived, leaving no before-and-after record of how classifications changed. FAS also notes that the Household Pulse Survey restored gender identity fields in historical data after a court order but no longer asks them going forward. Both cases require provenance at the release level, rather than a note that merely says the dataset came from a federal domain, as the methodology records.
The practical response is plain version control for data. Consumers can save retrieval timestamps, response headers, schemas and checksums beside each snapshot. They can test expected columns and row counts before replacing a known copy. When licensing and storage permit, preserving the raw response makes later analysis possible even after a public interface changes. These steps follow from the tracker team's finding that status can change after court orders, collection mergers or agency revisions, and that its own termination count is versioned and updated at least quarterly in the FAS account.
US data changes travel beyond US borders
One terminated program published pollution readings from US embassies around the world. The tracker says the State Department switched off that network in March 2025 over funding constraints and that about 17 years of readings were removed from AirNow. Those monitors offered an external reference in cities where local air-quality figures could be disputed. The loss therefore reaches researchers, journalists and residents outside the country that operated the sensors, as the tracker's embassy-air entry explains.
USAID's Demographic and Health Surveys have a similar reach. The tracker describes four decades of household surveys covering health, fertility and nutrition in more than 90 lower- and middle-income countries. Other funders have stepped in to complete current rounds and maintain historical access, but that does not make the transition invisible. An organization ingesting the series must record which institution produced each round and whether collection methods stayed comparable, rather than treating the old URL as the dataset's identity, according to the tracker record.
The same issue appears in recreated domestic products. NOAA stopped updating its Billion-Dollar Weather and Climate Disasters series after the 1980 to 2024 archive. Climate Central hired the lead scientist and continued the tally with the same stated methodology, the tracker says. That preserves useful work, though the publisher and stewardship have changed. FAS includes this type of derived product when recreating it requires unusual resources or when its intended use depends on government authorship, a judgment call set out in the methodology.
The tracker is evidence, with limits
The project does not claim perfect visibility. Federal agencies are not uniformly required to announce when a collection ends, and sensors or satellite products may leave little paperwork. The team monitors regulatory filings, agency sites, news reports and tips, then reviews entries line by line and deduplicates products. It also avoids using the raw data.gov catalog counter, which can swing when sites are reorganized and can count many files from one survey as separate datasets, according to FAS.
Those choices make the inventory more useful, but they retain judgment. The tracker can include a planned follow-up that did not happen, a product that stopped updating and a proposed termination whose final decision is still pending, each with a caveat. It can also revise a status when a court restores access. Readers should use the linked primary records for any consequential decision and cite the tracker with an access date. The Guardian's earlier 28-plus-338 snapshot and the current 375-entry page already show why a bare total ages quickly.
The next test is whether agencies provide durable change notices that machines can consume. Watch for the tracker's next quarterly version, restored collections after litigation, and new releases that silently omit established fields. If federal publishers add dated schemas, retirement notices and archived snapshots, downstream maintainers can handle policy changes as migrations. Until then, a green status check only proves that the server answered, while the versioned tracker may record a deeper break.