How records are collected and reviewed.

01 · Scope

Coverage

We track opposition to U.S. data-center development, including policy, permitting, zoning, moratoria, litigation, elections, public comment, demonstrations, project outcomes, and directly related public-safety developments. General industry news is out of scope unless it is relevant to one of these subjects.

02 · Sources

Source selection

We prefer official records and primary documents, followed by reputable reporting. A statement by an involved organization supports what that organization said; it is not independent confirmation. Aggregators and social posts are treated as leads.

03 · Verification

Verification and confidence

Records distinguish primary confirmation, independent corroboration, single-source reporting, unverified leads, disputed claims, and corrections. Claims with serious consequences receive additional review. Missing information is left blank rather than inferred.

New-event findings from scheduled daily searches may publish after automated source, claim, and duplicate checks. These records are labeled “Not analyst-reviewed” until a human analyst validates them. Possible duplicates, sensitive claims, and findings with insufficient evidence remain in the review queue.

04 · Machine-assembled developments

Machine-assembled developments and their accuracy

The analyst workspace also shows developments assembled without human review. Automated weekly searches across news indexes and the open web, state by state, collect reports; a language model keeps those relevant to the scope above, labels each with a state, type and date, and groups reports about the same real-world event into one development. Each such development is labeled “Machine-assembled · not analyst-reviewed” and lists every report it rests on, so it can be checked against its sources.

The estimated accuracy shown beside that label is the share of a random sample of released developments that a reviewer judged accurate: a real, in-scope event, placed in the right state, of the right type, and not a duplicate of another development in the sample. The sample size and date are shown with the estimate, which is measured again when the data changes materially. It is an average: an individual development can be wrong even when the estimate is high, and the labeller's confidence shown on each development is a separate, unvalidated signal.

Trends count developments, not articles. Each is counted once, on its event date when the reports give one inside the weeks it was found, otherwise on the day it was first found. A change in search coverage can move counts as much as a change in activity does.

05 · Limitations

Limitations

A completed search cannot establish that no relevant development occurred. Sources may be delayed, unavailable, removed, or hard to index. Daily event records are listed as they are published; weekly bulletins report the preceding week's publications and the number of days with complete collection. A quiet or incomplete week is not evidence that no development occurred.

06 · Corrections

Corrections and record history

A material correction records the date, reason, and supporting source. Earlier publication history is retained. Record merges and consequential changes require human approval during the initial operating period.