Method and sources

Version 1.0, November 10, 2026.

This is a desk review of public evidence about named projects. The record should help a resident or decision-maker identify a documented benefit, an unresolved impact, a binding condition, and the next record needed to answer a question.

What an answer means

Each assessment answers one published question for one identified project, phase, and time period. The question must be readable without knowing the site's internal schema.

Answer	Publication rule
Yes	The cited evidence directly supports the exact statement in the question. The accompanying evidence label explains who says so and what was checked.
No	A cited source directly establishes the contrary. Silence, an unsuccessful search, and a missing document are not sufficient.
Not known	Evidence was not found, is insufficiently specific, is materially disputed, is outdated for the asserted period, or has not been reviewed. Explain which reason applies.
Does not apply	A documented fact makes the question inapplicable to this project or stage. Explain it. Do not use this to hide an information gap.

A “Not known” answer is not a negative finding. A yes is not automatically an environmental benefit: “Is a public air permit available?” asks about access to a record, not whether the facility has zero emissions. “Does the disclosed design avoid evaporative cooling?” describes a particular design characteristic, not the total environmental effect of the project.

Questions containing “public,” “published,” or “available” require a defined evidence scope. If the review cannot establish whether another record exists, use “Not known” rather than claiming the record does not exist anywhere. The notes may say “not found in the sources reviewed on [date].”

Keep the answer and its evidence separate

Evidence label	What it establishes	What it does not establish
Company statement (`operator-claim` in the data)	A named developer, tenant, or operator made a project-specific statement.	Independent verification; an enforceable obligation; delivered results.
Official record (`official-record`)	A government or utility record documents a decision, filing, permit, tariff, or agreement.	That every assertion in the filing was independently checked. An applicant's estimate in an official docket remains the applicant's estimate.
Measured result (`operating-data`)	A source reports results for an identified operating period and boundary.	Independent assurance unless the assurance statement, scope, and provider are also available.
No mark (`not-reviewed`)	No evidentiary judgment has been made.	Either satisfactory or unsatisfactory performance.

For a larger system, add separate fields for author, recordHost, bindingStatus, measurementMethod, and assuranceStatus. Document hosting, binding effect, measurement, and independence are different properties; a single four-value label is only a readable summary. Retain mixed evidence where a public permit quotes an operator's design and separately imposes enforceable limits.

A regulator-approved permit can support “a limit was imposed.” Monitoring or inspection records are needed to support “the limit was met during the stated period.” EPA's ECHO system contains reported permits, inspections, violations, enforcement actions, and penalties, and identifies data completeness and refresh limitations. No matching ECHO result is not a certificate of compliance. EPA ECHO data description.

Do not produce a universal clean score

The site should not add these answers into a 0–100 rating, star rating, or best-to-worst league table. The questions mix disclosure, design, local legal obligations, and results. A project could publish extensive records and still impose substantial impacts. Another could have relatively low impacts but inadequate disclosure.

The useful comparisons are:

the same question, same project stage, same reporting boundary, and a comparable period;
filters for a particular documented feature or missing record;
separately displayed measured quantities with units, boundary, method, and uncertainty;
progress from commitment to adopted condition to measured result;
an explicit checklist selected by the reader, with every unanswered condition shown.

Do not make high disclosure counts a proxy for environmental quality. If an evidence-availability count is shown, call it a count, show its denominator, and do not rank projects by it. Any future “meets these conditions” label must enumerate a versioned, public conjunction of requirements. One missing required condition prevents the label; a stronger result in another category cannot compensate for it. That label would still not mean the project has no harms.

The first published questions

The initial project records use the following nine questions. Their versioned definitions live beside the dataset in data/audits.json and are displayed in the site.

ID	Published question
disclosure	Is a project-specific public disclosure or government project record available?
cooling	Does the disclosed design avoid evaporative cooling throughout the reviewed campus?
water	Is a numerical site or campus water demand or consumption figure public?
energy	Is project-specific clean-electricity accounting public with an explicit matching period?
ratepayer	Is an adopted tariff or agreement assigning project electricity costs publicly verified?
air	Is a final project-specific air permit publicly available in the reviewed records?
noise	Are enforceable numerical operational noise limits identified in the reviewed records?
decision	Is a local land-use decision documented in a public record?
performance	Is measured site or campus environmental performance publicly reported with a reporting period?

A yes to energy can describe annual matching; inspect the note rather than assuming hourly matching. A yes to water can describe a forecast or a measured value; the type, unit, period, and boundary remain essential. A yes to performance only covers the reported metric and period, not every environmental outcome.

These are core record questions. They do not constitute a full environmental, social, fiscal, or engineering audit. The expanded specification below identifies what would need to be collected before claiming comprehensive project assessment. Fields not yet collected must not be inferred from an operator's corporate sustainability report.

Expanded question specification

These questions are proposed collection requirements, not claims that the launch has already assessed them. Some need several independent yes/no fields to avoid a vague “responsible” judgment. Numeric observations remain numeric; the yes/no answer tests whether a precisely defined fact or condition exists.

Project identity, scope, and disclosure

Is the site boundary or named parcel set publicly documented?
Are the landowner, developer, facility operator, and disclosed tenants separately identified?
Does a project-specific public record identify the current application or permit number and phase?
Are the claimed building area, disturbed area, ultimate capacity, and initially requested capacity distinguished?
Are load, water, and job estimates accompanied by a date, phase, and author?
Are the source documents necessary to check the principal claims publicly accessible?

An undisclosed tenant stays undisclosed. A brand on a press release is not proof it owns the land, operates the cooling plant, controls the utility contract, or is the permit holder. Acquisitions and project renamings require dated relationships, not replacement of historical names.

Electricity, emissions, and reliability

Is site electricity consumption reported for a defined operating period, separately from planned MW capacity?
Are IT load, total facility load, utility reservation, and ultimate campus capacity distinguished?
Are both location-based and market-based electricity emissions reported with methods and reporting periods?
Is project-specific annual renewable or carbon-free procurement documented, including ownership and retirement of energy attributes?
Is hourly matching reported for the relevant grid region and actual project load?
Are the generation projects, commissioning dates, and contractual basis for claimed new supply disclosed?
Are battery charging sources, losses, and the treatment of stored electricity disclosed when batteries support a clean-energy claim?
Are actual PUE and its measurement boundary, period, and occupancy/load conditions published?
Are any claimed load flexibility, curtailment, and backup-power arrangements documented and, for an operating site, tested or measured?

Interpretation rules. MW is instantaneous capacity; MWh is energy over time. Do not multiply a headline MW figure by 8,760 hours and present the result as measured annual use. If a scenario estimate is useful, disclose utilization, PUE, calendar period, and range, and keep it outside the verified observations.

Annual certificates can support a properly bounded renewable procurement claim. They do not establish that the facility uses carbon-free electricity every hour or that a specific new generator would not otherwise have been built. EPA explains that RECs represent the environmental attributes of generated renewable electricity. The UN's 24/7 CFE definition concerns matching consumption every hour. EPA REC explanation, UN 24/7 Carbon-Free Energy Compact.

Use the applicable published accounting standard, and record its version. GHG Protocol's Scope 2 guidance distinguishes contractual accounting and specifies quality criteria; its current page also describes the 2025–26 revision consultation. A proposal under consultation is not an adopted requirement. GHG Protocol Scope 2 guidance.

Regional emissions factors can provide context or a labeled estimate. They are not a meter reading from the project. EPA recommends eGRID subregion output rates for estimating electricity-use emissions; do not substitute an average factor for a causal estimate of the project's marginal effect on the grid. EPA eGRID questions, EPA calculation guidance.

Water and cooling

Does the complete heat-rejection design avoid evaporative cooling, including peak-day or backup modes?
Are annual and peak-day water requirements quantified for each buildout phase?
Are withdrawal, delivery, consumption, return flow, and wastewater discharge distinguished?
Are potable, reclaimed, surface-water, groundwater, and other sources separately quantified?
Is an authority or utility assessment of available capacity and drought conditions public?
Are water allocation, drought curtailment, discharge, and monitoring conditions documented where applicable?
Are site WUE, absolute water quantities, and the calculation boundary reported for the same operating period?
If replenishment is claimed, are project location, timing, additionality, method, verification, and beneficiary watershed disclosed?

“Closed-loop” or “liquid-cooled” is insufficient to answer question 16. DOE illustrates designs in which a recirculating liquid loop ultimately rejects heat through an evaporative cooling tower. Examine the full heat path. PUE measures facility energy relative to IT energy; WUE relates site water to IT energy. Neither substitutes for absolute demand or local context. DOE cooling and efficiency guide.

Withdrawal and consumption answer different questions. A large withdrawal can constrain pipe or source capacity even if some water returns; consumption concerns water not locally available for reuse. Preserve the source's definition rather than relabeling all reported “water use” as consumption. USGS water-use terminology.

Keep direct facility demand, electricity-related water, and supply-chain water separate. Replenishment elsewhere does not automatically repair a local peak-day shortage. National averages cannot establish a particular project's water effect; local source, season, system capacity, and other users matter. Berkeley Lab's national report is useful context for separating direct and indirect demand, not a facility audit. Berkeley Lab 2024 report.

Local air, fuel, and health context

Is an applicable air permit or documented exemption available, with a matched facility and equipment inventory?
Does the record distinguish emergency standby, routine testing, demand response, and continuous on-site generation?
Are permitted pollutant quantities, operating-hour limits, stack parameters, fuel, and controls public?
Is the applicable air-quality analysis public, including nearby receptors and cumulative sources where evaluated?
Are actual runtime, emissions, compliance evaluations, and unresolved enforcement findings available for a defined period?
Are fuel storage, spill prevention, fire response, and battery safety conditions documented where applicable?

Do not equate permitted potential emissions with actual emissions, or emergency equipment nameplate capacity with continuous operation. Do not equate the presence of a permit with absence of health effects. State permit requirements depend on equipment and scope. Minnesota's regulator explicitly describes air, water, stormwater, and fuel-storage responsibilities and enforceable monitoring conditions. Minnesota Pollution Control Agency.

Noise, land, construction, and neighborhood effects

Is a site-specific acoustic study public, with operating and testing scenarios, receptor locations, and background conditions?
Are enforceable noise limits stated with time period, measurement location, units, and any tonal or low-frequency provisions?
Are post-commissioning measurements and a complaint investigation process documented?
Are residential setbacks, lighting limits, visual screening, and construction hours specified in an applicable decision?
Are wetland, stream, floodplain, habitat, cultural-resource, and land-disturbance impacts documented?
Is the relevant avoidance, mitigation, and monitoring obligation public, separately from voluntary landscaping claims?
Are roads, truck traffic, stormwater, wastewater, and emergency-service impacts and funding responsibilities documented?

A national dBA cutoff would obscure different local rules, measurement methods, and types of noise. Store the adopted local condition and measured result; commission a comparable protocol before ranking sound levels across projects. Fairfax's adopted data-center zoning amendment is an example of local siting and design conditions, not a rule for every American municipality. Fairfax adopted amendment.

For wetland work, map screening is a lead, not a legal jurisdictional determination. Section 404 covers certain discharges into jurisdictional waters; permitting bodies and exemptions vary. Read the actual decision and current state requirements before claiming a project needed, received, or violated a permit. EPA Section 404 program.

Ratepayers, taxes, employment, and public benefits

Is the applicable large-load tariff or approved service agreement public and actually applicable to this customer?
Are responsibility for generation, transmission, distribution, interconnection, and stranded-asset costs separately documented?
Are minimum payment, contract length, collateral, exit charges, and cancellation obligations enforceable and disclosed?
Are tax abatements, exemptions, grants, public borrowing, and infrastructure subsidies quantified with their beneficiaries and term?
Is a public net fiscal analysis available with assumptions, uncertainty, and service costs?
Are permanent direct employees, contractors, construction jobs, job-years, and indirect/induced estimates distinguished?
Are wage, local hiring, training, and community-benefit commitments enforceable, measurable, and accompanied by reports?
Is a decommissioning or reclamation obligation and its financial assurance documented?

An infrastructure contribution is not proof that every ratepayer risk is eliminated. An adopted tariff is stronger evidence than a press statement but must be read for applicability and exceptions. The Ohio Consumers' Counsel records adoption of AEP's data-center tariff and subsequent appeal; that source does not establish the terms for a different utility or state. Ohio case overview and docket links.

Jobs and tax benefits deserve the same evidentiary discipline as environmental impacts. Virginia's legislative audit distinguishes substantial construction benefits from ongoing employment and describes variation in local tax revenue. Do not turn an economic-impact multiplier into a count of direct hires, or gross economic output into municipal revenue. Virginia JLARC report.

Community process, distribution of effects, and continuing accountability

Are the responsible decision bodies, current decision stage, next public meeting, and comment deadline documented?
Were the relevant notices, staff reports, testimony, responses, and final vote publicly recorded?
Are benefits and burdens described at the affected neighborhood or watershed level, with a stated method?
Are affected Tribal governments and other legally required consultation parties identified in the relevant record?
Is there an accessible complaint and correction route, and are material complaints linked to findings rather than treated as findings?
Are changed designs, amendments, ownership changes, incidents, and expired commitments reflected in the record history?

A hearing shows a process occurred; it does not prove unanimous consent. A petition shows expressed opposition; it does not prove an environmental violation. Demographic and cumulative-burden data are context. They do not by themselves establish that a particular project caused a health outcome or acted unlawfully.

Materials and lifecycle

Is a project-specific embodied-carbon assessment available with system boundary and quantities?
Are refrigerants, leaks, equipment replacement, and construction materials included in the claimed emissions scope?
Is a documented reuse and recycling chain available for servers, batteries, and other equipment?
Is recycler certification current for the named facility and the relevant service scope?
Is any claimed waste-heat reuse supported by an identified receiving customer, useful delivered heat, and a measured period?

A corporate recycling policy is not a project-level chain of custody. EPA describes R2 and e-Stewards certification as independent recycler standards; certification of a recycler does not certify the environmental performance of a data center that names it. EPA certified electronics recyclers.

Port Washington: how to read the example

Oracle's project page is more specific than a generic corporate sustainability statement. It describes cooling, an energy commitment, land area, infrastructure funding, and expected jobs. It also dates its information to January 2026 and characterizes future statements as forward-looking. That makes it useful source evidence, while leaving engineering delivery, binding conditions, and actual outcomes to other records. Oracle Port Washington disclosure.

A correct record can therefore recognize a disclosed non-evaporative design while leaving measured water demand unknown. It can report a clean-energy commitment while asking about hourly matching, accounting boundaries, new supply, and performance. Positive features and unanswered questions can coexist. The detailed project record supplies the actual cited answers; this section does not independently certify the project.

Publishing and correction rules

Preserve the exact source URL, title, publisher, source date, retrieval date, relevant section/page, and a concise paraphrase. Keep source snapshots and hashes privately or publicly only where reuse rights allow.
Separate the source's effective date from the date it was checked. A page fetched today may repeat a two-year-old plan.
Assess a specific phase. Existing buildings may operate while an expansion is only proposed.
Require direct support for every yes and no. Cite the contradictory material when sources conflict. Do not resolve a material conflict by choosing the more favorable or newer-looking page without checking its scope.
Have a second reviewer check adverse findings, identity matches, numerical conversions, and claims of legal obligation before publication. An automated extraction is a draft, never the approving reviewer.
Publish a concise revision history. Corrections should identify what changed and why; do not silently erase prior material findings.
Provide a correction and right-of-reply channel for residents, developers, operators, authorities, and other affected parties. Request the project, disputed statement, supporting record, and preferred contact. Do not publish personal contact information with the response. Data Center Evidence has no online form. Corrections and replies reach the operator by post, at the mailing address at the end of the Privacy page.
Link a substantive reply to the disputed issue. Payment, advertising, or cooperation cannot change an assessment standard. An unanswered request is not evidence of a violation.
Use descriptions of conduct and evidence, not labels about character: “the public permit allows X,” “the operator reports Y,” or “the reviewed record does not resolve Z.” Do not publish “clean,” “dirty,” “good actor,” or “bad actor” as an unqualified verdict.
Reassess when a decision changes, a new report arrives, or the freshness interval expires. Show overdue review explicitly; do not fabricate continuing monitoring.

These are operating requirements for a comprehensive service. The initial published dataset must describe the review actually performed and must not imply that a second human review, outreach, or field measurement occurred when it did not.

Where the map data comes from

The repository contains 2,219 data-center records from the openly licensed Compute Atlas v1.35.1 release. The preserved source is pinned to commit 85cd69d93a7a4e194f3a4d7e3723a4caebd88dde; its SHA-256 is recorded in data/inventory.json. The import excludes 226 crypto-mining and 478 dedicated power-generation records. It does not copy their environmental claims into the assessments.

Source-reported lifecycle	Imported records
Proposed	602
Permitted	150
Under construction	262
Operational	1,066
Cancelled	139
Total	2,219

These are source-defined sites, campuses and some multi-site records, not 2,219 independently verified buildings. There are records in all 50 states, DC, Guam, the Northern Mariana Islands, Puerto Rico and the US Virgin Islands. Nine records are in territories. American Samoa has no record in this snapshot; that is not evidence of no data centers there. Record update dates range from July 16 through October 31, 2026. Retrieval on November 10 does not make those statuses current on November 10.

Of these locations, the source labels 1,820 approximate, 394 exact and five representative of multiple sites. Those labels are preserved; the coordinates are not independently verified. The source calls 833 records confirmed, 1,359 reported and 27 rumored. These are the source's confidence labels, not audit decisions. Keep rumored records visible as leads if useful, but do not present them as verified projects.

Data credit: Compute Atlas by Edward Kubiak, CC BY 4.0. Changes: filtered to data centers and normalized discovery fields. The pinned data license explicitly permits sharing and adaptation with attribution. Compute Atlas does not endorse the separate assessments. The preserved source retains original source links, notes and lifecycle histories for later reviews. The normalized file includes source citations without asserting that each one has been read.

Reproduction:

python3 scripts/inventory.py
python3 scripts/inventory.py --fetch
python3 scripts/inventory_sources.py

The first command rebuilds deterministically from the preserved gzip; the second fetches the same public commit and rejects an unexpected checksum. It does not use a gated current export. The third rebuilds a 56-jurisdiction worklist from EPA's official agency directory. The worklist states that permit monitoring is not yet implemented; it is a roster of official starting points, not 56 functioning crawlers.

National agency worklist

data/inventory-jurisdictions.json contains the official agency starting points extracted from EPA for 50 states, DC and five territories, along with current imported-record counts. Each entry explicitly marks permit monitoring and systematic local-planning review as unfinished. Inventory coverage reaches 55 of those jurisdictions; the source has no American Samoa records. Air districts, tribes, utility commissions and local planning authorities still need their own child entries and checks.

Before describing a state as covered, require a documented source roster, named maintainer, successful recent retrievals, a reviewed backfill interval, identity reconciliation and a sample-based miss check. A state with 150 imported records can be less completely covered than one with 15; counts alone cannot answer that question.