How the system is built

Scope decisions

The system is designed for one question: how federal and selected state policy changes the environmental, economic, and political conditions for U.S. data center development. It favors depth, traceable sources, and manageable noise over an attempt to collect every government record that might mention technology or infrastructure.

• Federal first. Congress, federal rules and notices, executive actions, hearings, and federal research form the core policy record.
• Selected state context. State bills and governor actions are included when they pass data center specific discovery rules. This is useful context, not a complete fifty state legislative or regulatory service.
• Public sources first. Records should lead back to an inspectable government page or clearly identified external source.
• Deliberate exclusions. The system does not claim comprehensive local zoning, parcel level permitting, PUC/PSC, ISO/RTO, or FERC eLibrary coverage. The local action lane names its selected markets and included signal types; the deadline feed is limited to observed federal comment dates and tracked congressional meetings.

How records enter the system

1. Discover. Scheduled collectors check defined public sources. Each source has its own bounded search: recent Congress records, Federal Register searches scoped by keyword and agency, title gated state legislation, official governor pages, and selected federal research feeds or registries.
2. Normalize. Source specific fields are mapped into a consistent record with an identifier, title, dates, status, jurisdiction, source URL, and available sponsor or agency details. The original source remains linked.
3. Filter and classify. Source rules remove obvious noise. A language model then assigns relevance and one or more topics from a fixed vocabulary. Items below the default relevance threshold remain available through the low relevance filter rather than being silently deleted.
4. Enrich. Deterministic text rules identify possible policy mechanisms such as utility rates, water standards, tax incentives, disclosure, siting, or interconnection. These are research signals, not legal conclusions.
5. Track change. A content fingerprint is compared with the stored version. Material status or stage changes are logged, and changed records are queued for search indexing and downstream briefings.
6. Present with provenance. The interface keeps official policy records, external facility observations, voluntary guidance, contextual research, and model synthesis in separate evidence lanes.

Source and inclusion choices

Congress.gov

Recently updated bills are scanned and retained when their titles match defined data center terms. Relevant committee meetings are kept when the meeting title matches or it links to an already tracked bill.

Federal Register

Full text data center searches are limited to selected agencies, with a separate keyword gated pass for presidential documents. Regulations.gov adds docket level comment context to matched records.

State legislation

State discovery combines selective official sources and legislative data providers. Most searches use data-center or large-load terms. Ohio also checks selected electricity, water-access, industrial-permitting and infrastructure-tax titles, then verifies bill identity, session, dated history and current text links. These wider candidates still need relevance review; broad bills and unexpected wording can be missed.

Executive actions

Official governor sources for all states and D.C. are monitored source by source. Failed or blocked checks are reported; they are not treated as successful zero result searches.

Research and context

GAO, GovInfo, selected CRS material, grid data, playbooks, facility references, and contextual sources help explain policy. They do not establish law unless the cited government record does so.

How the spatial and official action maps work

• Facility Atlas. The map imports three legally reusable public datasets: Compute Atlas ↗ (CC BY 4.0), PNNL's IM3 Data Center Atlas ↗ (OpenStreetMap derived locations under ODbL), and Epoch AI Data Centers ↗ (CC BY 4.0). Each file is structurally validated and checksum receipted. Source identities remain separate: an IM3 row is coverage, not independent corroboration of another OSM derived observation. Epoch's power fields remain labelled as estimates, and its published U.S. addresses are mapped through the Census batch geocoder as approximate context. God's Eye View contributes no facility rows.
• Map imagery. Street tiles come from OpenStreetMap. Optional aerial context comes from USGS The National Map orthoimagery. An aerial image is centered on the published coordinate; it is not a verified photograph of a facility, and approximate or representative coordinates may show only the surrounding area.
• Cross source identity. A bounded generator proposes same state candidates using geographic distance plus name and operator token overlap. It cannot merge records. Confirmation requires an authenticated reviewer, note, date, and optional evidence link; both observations remain separately attributed.
• Verified facility media. A building photograph appears only after staff stores its exact observation match, publisher page, usage right, accessible description, capture precision, match basis, and review receipt. Aerial tiles and renderings are never promoted into that category.
• Context layers. Incentive, permitting, moratorium, and grid layers count source linked state policy activity. EIA measurements remain state or balancing authority context. The USGS Surface Water Supply and Use Index layer uses historical 2019 model output aggregated from HUC12 watersheds to states; it excludes groundwater and is not a site determination.
• Documented official actions. The position ledger supports sponsored measures, votes, official statements, and executive actions. Every item keeps its date, role, recorded direction, evidence status, and underlying link. Changes remain in the timeline rather than overwriting earlier evidence. It describes actions, not a permanent political stance.
• Geographic meaning. A federal legislator maps to the state represented. A state measure or executive office maps to the record jurisdiction. State shading counts evidence items and is not a sentiment or influence score.
• Conservative categories. Explicit title evidence takes precedence over a conflicting derived mechanism tag. Mechanism categories remain navigation aids and must be checked against the linked record.

What automation does

Automation collects and normalizes records, detects changes, assigns relevance and topics, identifies mechanism signals, indexes source passages, and drafts research answers or briefing language. Fixed identifiers, source labels, change logs, citations, and record relationships use deterministic rules wherever possible. AI helps with ranking and synthesis; it does not determine legal applicability or replace the source document.

Assumptions and evidence rules

• A source's published status and dates are treated as the best available record until that source changes or a correction is documented.
• Sharing a state, subject, or actor makes two records candidates for comparison. It does not prove that one caused, governs, or legally applies to the other.
• A direct relationship is shown only when inspectable evidence explicitly supports it. Same jurisdiction policy is otherwise labeled as context.
• A missing search result means the maintained corpus did not support an answer. It does not prove that no relevant policy, filing, or local activity exists.
• Government requirements, voluntary guidance, publisher claims, contextual journalism, and model generated suggestions remain visibly distinct.

How quality is checked

• Source receipts. Scheduled jobs record whether a source check succeeded, found no matches, or failed.
• Freshness and backlog checks. The data health audit checks source recency, classification queues, missing summaries, mechanism coverage, and briefing freshness.
• Retrieval evaluation. A maintained Policy Desk question set tests geography, identifiers, sponsors, dates, mechanisms, comparisons, recent changes, and evidence boundaries.
• Briefing review. Automated checks test structure, dates, source links, duplication, and evidence. A briefing counts as substantively reviewed only when a human review is logged.
• Application tests. Automated tests check data contracts, navigation, evidence labeling, and critical routes before a change is released. Production smoke checks confirm that public pages and core research paths still respond.

The live source registry and corpus receipt expose configured scope, cadence, latest receipts, known boundaries, and a reproducible snapshot identifier. The machine readable manifest is available at /api/corpus/manifest ↗.

Completeness and uncertainty

This is a maintained research corpus, not a complete census of every policy action. Keyword choices, title filters, agency scope, publication delays, inaccessible government pages, source metadata errors, and classification mistakes can all create gaps. The system is strongest when it helps a researcher find and compare inspectable evidence; it should not be used as proof that no other relevant activity exists.

Refresh timing varies by source, so the public record may update before this system does. Current source status and processing gaps are published on Data health. More specific exclusions are listed under Limitations.

Corrections and further detail

Classification and summaries can be challenged. See AI use and corrections for where models are used and how to report a problem, or review Validation status for current evaluation results and open launch gates.