Methodology

How the US Data Center Opposition Tracker is built

Author

Jaewook Lee

The tracker is a source-verified relational dataset of organized opposition to data center installation in the United States. It covers local, state, and national campaigns in one schema. Release 1.2.1 contains 265 movements, 670 events, 780 actors, and 476 sources through September 30, 2026.

This page summarizes how the tracker is built. The project README and the citation are in the published repository.

The verification rule

Every release applies the same rule:

No factual claim enters the tracker without a supporting source URL and an access date. Aggregators and search results are leads, not final evidence: entries are checked against the underlying news report or official record. If a source cannot be opened, checked, or manually verified, it is treated as nonexistent and removed completely.

Each event in the dataset is linked to at least one source with a working URL. Publication and access dates are recorded as they occurred. When an automated capture of a source fails, the page is checked by hand in a browser before the source is kept.

Four tiers of discovery

The search is built to find failed moratoriums, approved projects, and unresolved campaigns as well as opposition wins. Each tier below is illustrated with a case added in the September 2026 update.

Tier 1: Structured inventories and reports

Research reports, maps, and newsletters point to cases, and the tracker cites the official record or local reporting behind each one.

Example: a July 2026 newsletter item led to Friends of Hanover’s campaign against Tract’s Mountain Road Technology Park in Hanover County, Virginia. The tracker cites VPM’s original report, which dates the supervisors’ 4–3 rejection of the project to May 27, 2026, two months before the newsletter.

Tier 2: Official legislation and government records

Each state, the District of Columbia, and the federal government gets one search of a bill tracker for data center bills on moratoriums, siting, and tax incentives. Municipal and county agendas, zoning ordinances, court dockets, and agency releases are searched when a lead points to them.

Example: a legislative search for North Dakota led to the state’s 2026 special session. Legislative Management declined to advance a data center moratorium on August 17, and SB2406, a narrower bill on nondisclosure agreements, failed on Senate second reading, 15–32, on September 3. The movement is recorded as lost, with the legislature’s bill overview and action history as sources.

Tier 3: News searches across 52 jurisdictions

Each state, the District of Columbia, and the federal government also gets one news search for recent reports of data center opposition, moratoriums, and protests. Directory and aggregator sites are left out so that original local reporting comes first. Every search and its results are logged, including searches that found no new case.

Example: the Iowa news search surfaced Iowa Public Radio’s report that the Salix City Council rejected a one-year moratorium, 3–2, on August 20, 2026, despite the Salix Guardian Alliance’s call for more study. The lost status applies to that moratorium request; rezoning and annexation litigation over the project were still pending.

Tier 4: Snowball discovery

Names of towns, groups, developers, bills, appeals, and related projects in each verified source are followed up.

Example: after verifying Eastport, Maine’s 180-day moratorium on land and underwater data centers (August 12, 2026), following up on the case led to the Center for Biological Diversity’s September 24 move to intervene in federal permitting, citing risks to endangered whales. The intervention was added as a legal_action event.

The six-phase update workflow

Each release runs through six phases. The reusable runbook is EXPANSION_GUIDE.md in the research repository.

  • Phase 1 resolves the highest-priority open questions from the previous round before any broad search.
  • Phase 2 reviews every existing movement, with review depth recorded as source_verified, search_only, carried_forward, or unresolved.
  • Phase 3 adds missing movements through systematic discovery across the four tiers above.
  • Phase 4 replays each phase from its saved baseline and audits consistency across the canonical tables.
  • Phase 5 archives the sources added since the last release, rebuilds derived outputs, and builds the local release candidate.
  • Phase 6 writes the public update note, exports the strict public allowlist, verifies the live site, and records the publication state.

AI tooling

I used AI coding assistants throughout the project: Anthropic’s Claude Fable 5, Opus 4.6, Opus 4.7, and Opus 5.5, through Claude Code. They wrote the Python and R scripts and drafted documentation. In the September 2026 update, an AI session also ran the web searches, recording every query and its results, and read the retrieved pages to draft candidate events. The scripts themselves do not call a language model, so the same inputs produce the same outputs. I check every candidate event against its underlying source before release, including any page that has to be opened by hand because automated retrieval failed. A source that still cannot be verified is removed, along with the claims it supported.

The search process has changed as the tracker grew, from state-by-state sweeps in the first two releases to fixed query patterns in the September 2026 update.

Reproducibility

Every data change belongs to a numbered batch in batch_log.csv with exact before-and-after values. Phase 2–4 integration scripts stop if the saved baseline no longer matches the live tables. Downstream research projects use frozen snapshots exported by scripts/export_snapshot.py. Releases follow semantic versioning and are cited through CITATION.cff.

For more detail

A longer methodology document covers the status taxonomy, the privacy and publication boundary, a worked example (Nottingham, New Hampshire), and the current limitations. It is maintained with the research tables. To request it, email to.jaewook.lee@gmail.com.