How the Record is built.

A record carries exactly one of nine statuses, and every change of status is a dated, sourced entry in its history. The definitions:

• Reported. Named in reporting, with no public filing found. A codename, a land assembly, a utility inquiry, not confirmed by the developer, a landowner or a public body. The data status behind it is rumored, which is what every export and every query carries; the page says reported because that is what the record knows, and calling it rumored reads as though the record doubted the reporting.
• Proposed. The developer or a public body has confirmed the project and no application for it is on file. A project waiting on an enabling ordinance or text amendment before it can apply is proposed, however many hearings that ordinance takes.
• Permitting (shown to readers as waiting on permits). An application in the project's name is on file with a public body: a rezoning, a conditional use permit, a site plan, a water right, an air permit, an interconnection request. Permitting begins the day the application is filed and lasts until the last approval the project depends on is decided.
• Approved. The last discretionary approval the project needs has been granted and no approval it depends on is still pending.
• Under construction (shown as being built). Site work or building construction has begun, on the record of a permit, an inspection, a groundbreaking, or reporting from the site.
• Operational (shown as running). The facility, or its first phase, is serving load.
• Stalled. The developer or a public body has said the project is paused, or no action has been recorded for a year while approvals remain pending.
• Cancelled. The developer or a public body has said the project will not proceed, or its approvals have lapsed or been revoked.
• Not yet verified. A public source shows this site exists and names no stage for it; its current status has not been confirmed from a record. A record the EPA's facility registry brought to the record carries it until a source states its stage, because a permit says a facility is registered, not where a project stands: the generators of a data center running for a decade hold the same air permit as those of one still in the ground. So does a record whose only source is an operator's list that calls the site in consideration or planned, words that name no stage; a list that names a site among those its operator runs, or calls it under construction or coming soon, states running or being built, and the Pennsylvania DEP's tracker, a register of proposed projects, states proposed. It has a color of its own on the map and a line of its own under the count, and it is never counted with the four stages. On every page the count of records is the verified records, and a site not yet verified is named apart from that count, never inside it. The data status behind it is not_verified.

The line the Record draws is the application. A project waiting on an enabling ordinance is proposed; permitting starts when an application is on file. A status is never edited in place: when the rule changes how a project reads, a new dated entry is added to its history citing the source it rests on, and the earlier entries stay visible.

The front page says what the record holds in one line a section; the rules behind each line are here (October 9, 2026).

• How many. A record is one data center site: where it is, what stage it has reached, what has been disclosed about it, and the source for every line. Every number on the front page is counted from the published records when the page is built, and every item links to its source. The four stage words and the eight statuses behind them are under What each status means.
• Not yet verified. A site the EPA's facility registry places, or one an operator lists only as in consideration or planned, is on the map and named apart from the count: a public source shows the site exists and names no stage, and its current status has not been confirmed from a record. It is in none of the front page's numbers. The status.
• Projects. Proposals, approvals, construction and openings, from the first filing to the day a site serves load. A history is never overwritten: a change of status is a new dated, sourced entry.
• Power and water. Grid connections, on site generation and water commitments come from utility filings and permits, each figure with the basis its source gave. The count of records that disclose a figure is over the published records; the rest have not disclosed either figure in any filing the record holds, and the record never estimates one. The figures and the arithmetic.
• Jurisdictions. How states, counties and cities handle data centers: the votes, ordinances, moratoriums and tax rules, each row rated from open to closed by what the body has decided, under the scorecard rule.
• What people said. Statements by officials, executives, advocates and residents, recorded as said, dated and sourced, with no verdict; who spoke says how they are grouped.
• What the record covers. The record watches every US state, and counts a state as covered once one published record sits in it. Where the record stands today gives the live count and says how depth varies.
• The standards. Filings, dockets, meeting records and official disclosures outrank secondhand reporting, and every source shows its tier. A figure that is not in the public record reads not disclosed; when someone else estimated, the record says who. Every fact links to where it came from, so the record can be checked line by line. The full list is at the top of this page.

The Record's unit is the site: one operator building out one place, whether that place holds a single building or a campus of a dozen. An industry directory typically counts buildings, and a jurisdiction record (a permit, an ordinance, a scorecard rating) attaches to the site the buildings sit on, not to each building in it. So a directory's count and this Record's count answer different questions and are not directly comparable: a higher directory count does not mean more sites, only more buildings on the sites already here. Where an operator's own page names individual buildings on a site, those names are kept and shown on the record, and the count of them is published separately from the count of sites.

Coordinates come from sources that state where the site is: a county parcel record, a street address a filing gives, or the point the EPA's facility registry gives a permitted site. Every record carries how precisely it is located, in one of six words. Parcel, address and permit point mean the point is the site as a source states it. City, county and state mean the site is not disclosed and the point is that place's center, or for a county its seat.

A precise location is never inferred. It comes only from a source that states it, and that source is on the record's own source list. A town, county or state center stands in for a site the public record does not give: it is drawn with a dashed edge on the map, the map's facts panel says which center it is ("Location is the county center; exact site not disclosed."), and it is never presented as a located site, for the same reason no value is ever estimated.

The EPA's facility registries, ECHO and the Facility Registry Service it is built on, list sites that hold air, water or hazardous waste permits, with a street address and a point, and many data centers are among them. The record takes that point as a permit point only where the EPA says it placed the point at the site itself, by matching the street address, a survey or a photograph. A point the EPA placed at the center of a ZIP code, or whose placement it does not state, is not a site, and the record does not take it. The registry's facility page is cited as a primary source, with its registry number and the permit programs it lists. A location from the registry reaches a record only through an editor's approval. A permit never sets or changes a record's status, and a permit date is recorded only as the registry states it: it states when a permit expires, not when it was issued, so no permit issue date comes from it.

Where the record names a parcel, its outline is drawn on the record page from the county parcel record as published through the county or state parcel service, with the service and retrieval date shown. The imagery beneath the site map is the U.S. Geological Survey's national orthoimagery, a public domain federal work. No OpenStreetMap feature is ever copied into a record.

Every entry carries the date of the thing it describes, the vote, the filing, the publication, not the date the Record wrote it down. The Daily Record is the exception on purpose: it lists entries under the day they were written.

Some sources give only a month. An issue brief is dated September 2026 and a commission's assessment says March 2026, with no day on the page. When that happens the item, and any timeline entry, scorecard event or figure it produces, is dated the first of that month, and the entry says that the source gave the month only. The first of the month is a filing convention, not a claim that anything happened that day, so such a date prints as the month alone, "November 2025", on every page and in every export, where it reads 2025-11; on a timeline it sits in its month with no day given, and on the Daily Record it sits on the first of the month's page under the month. A year alone dates the first of January the same way. When the day is later found in the document or in a second source, a new entry carries it; the earlier entry is never edited.

A video is dated by its upload date, and that is the date of the source, not of the event. The event's date comes from what the video says, or the first of the month when it names only a month, never from the day the video was uploaded. An entry that was dated by an upload date is corrected the same way as any other: a new entry carries the date from the content and the earlier entry says so.

A third basis is a read date. Some pages carry no date at all: an operator's own list of its data center locations says the site exists, without saying when it opened. An entry from such a page is dated the day we read it, reads "Read on" and its date, and says only that the page named the site that day. A read date is never treated as an event: it does not put the record on the Daily Record or on a timeline, and it is never the newest record to go live or to break ground on a state page. It does stand on the record page as the entry it is, and it does set the record's status, because the page says the site is there.

A state page's summary card names the newest record in that state to go live, or if none, the newest to break ground, and only from a status change with a sourced event date. Where the state has neither, the card says "No dated groundbreaking or opening on the record" rather than naming a record at some other stage. Most states read that way today: the operator census that put hundreds of running sites on the record is a read dated citation, not a report of an opening, and the date an opening happened is usually not on the page that lists it.

Every timeline entry, statement and reference figure records where its date came from. A content date is the date of the thing described, taken from the source's own words. A source date is the source's own date, the day an article was published or a video uploaded, used when the content gives no date of its own; such an entry reads "Reported" and its date in Reader mode, and the Researcher toggle shows the basis on every entry. A planned date is a target the source states, later than the source's own date, a groundbreaking "next June" or an opening "in 2028"; it reads "Expected June 2025, as of November 2024", the second date being the source's. A statement made in the source itself, a person speaking in a video, is dated by the source; a statement the source reports from elsewhere is dated by the content.

How dates are sourced

Coming up on the front page, and What happens next on a state, county or project page, list hearings, votes, court dates and deadlines a source has stated. A date appears there only when a source gives it, and the body's own agenda or docket takes the place of a news report once it is on the record. Each item reads as one line, who acts, what and when, written from the source's own sentence and checked against it, with the source named beside it; where no written line passes that check, the line says only the kind of act and the day. Nothing is computed: a six month pause produces no end date unless a source states one.

Where a line sits on a timeline

A timeline shows what happened, on the day it happened. The rows it reads were not all written that way: most scorecard events carry the date of the article that reported them, and many lines name a time of their own that is not the row's date ("the town council voted 4 to 1 on August 6", filed on August 24). So each line is read from its own words before it is placed, and the row itself is never changed.

• One rule for every timeline, the national one, a year's, a state's, a county's or a city's, and a record's own: an act sits on the day it happened when the record knows that day, and an act known only by the date of the article that reported it sits under that article's month, one tap away, marked "Reported" with the article's date.
• A line that reports an act (a vote, an order, a filing, a ruling, a groundbreaking) sits on the day its own sentence names; failing that, on the weekday it names, counted back from the day of the report ("on Monday", in an article of Tuesday, September 22, is September 21); failing that, on the date its source gives as the day it happened. Where the only date is the article's, the act sits under the month as "Reported".
• A stage change sits on the day the record's status moved, "Reported" in front where that day is the article's, because that is the day the record changed.
• A line whose sentence names only a month or a year sits in that month or that year, with "no day given" or "no month given", never on a day it does not name.
• A time named as a bound or a target ("through March", "until November 2027", "by Monday, Oct. 19", "effective September 1") or inside a clause about something else ("the exemption, which was enacted in 2020") is not the time of the act.
• A line that describes rather than reports (what a law requires, what a site covers, a plan, a schedule, a statement) is not an event. It stays on the record under its month, one tap away, on its record page and on its state's timeline.
• The same act reported by several sources, about the same place or record and within three days, is one line with the other reports behind it, and the count of reports is shown. Where one of the reports dates the act by its own words, the act sits on that day with every report behind it; where none does, the reports are one line under the month.

The Researcher toggle shows, on every line, which of these placed it and the date the row was filed under. The reading is done by fixed rules over the line's words, not by a model, and each rule is a test case built from a published line.

The site wide timeline is a milestones index

Every event has one home record, and its record page is where its timeline lives: a project's event on the project page, a government act on the page of its state, county or city, a bill's stage on the bill page, a law on its laws page. The site wide timeline is a thin index of milestones that points at those pages, one month a page, and never a second copy of everything. A milestone is an act of one of these kinds: an approval, a denial, a moratorium, a groundbreaking, an opening, an expansion, a cancellation, a lawsuit filed, a court ruling, a law signed or taking effect, a federal roll call. It sits on the day it happened, read from its own words as above; an act known only by the date of the article that reported it is not on the index, and neither is an act of another kind, a filing, a statement, a rate rule, background or a plan, all of which stay on their home record and on the state's timeline. A line the record dates by the day a page was read is never an event on any timeline. The Daily Record and the timeline have different jobs: the Daily Record is the everything feed, every row by the day it entered the record; the timeline is milestones only, by the day they happened.

The yearly stage chart, which counted the projects on the record at the end of each year, is retired until the record holds a year of dated stage changes. The record began compiling in September 2026, and 541 of its 560 projects that day carried a stage dated by a report or a census read, so the chart showed when the record found a site, not when it was built.

A scorecard row carries one of four words: open, case by case, restrictive or closed, each defined by the kinds of events that earn it in the rubric below. The word is a person's. A row's rating is set only when a person accepts a rating in the admin, and it stays until a person accepts another: an event landing on the row never moves its rating, and a row with no rating stays unrated however many events it carries. Every event carries its date and its source, and the row's history is the list of them; an event may also carry the rating the classifier proposed for it, which is the proposal and not the row's word.

A rating is never set by a model on its own. The classifier that reads a source proposes one alongside the event it drafts, and that proposal reaches the row only where a person has accepted the word: by typing it on the card, by choosing it on the card's detail form, or by taking the proposed word with the Accept control, which names the word it is accepting before it does. Where nobody accepts it, the event still publishes with its own date and source and the row keeps the rating it already carried, so the scorecard does not move on a proposal alone; a row that has no rating yet takes the event with no rating, which leaves the row unrated and puts the fact on the record with its date and its source all the same. Undoing an approval does not put an earlier rating back, because the person's word stands until a person gives another. A proposal that is the word the row already carries is not a proposal at all, and nothing is asked about it. ratingPlan() is the one definition of what an approval writes, and the database function that files an event on a row touches the row's last verified date alone. Ratings that reached the scorecard before the acceptance rule, which took effect on September 21, 2026, were read against the rubric on September 29, 2026, as the changes below record.

A row lists what the body did, one line per act. A sentence with no act in it, background from the same article, a deadline, a plan or what a law says, is not an event and is not on the scorecard, though the row stays on file and on the Daily Record. Repeat reports of one act fold into one line: two sentences about the same row that report the same act within three days of each other, or a report that names no day within forty five days after the act it reports, are one line, the same matcher the timeline uses. The line's date is the day the best dated sentence names, else the month or the year where that is all a sentence gives, else the report's own day labeled Reported. The source that carries the line is the body's own page where one reported it (the governor's office for a directive), else the best dated report; every other report is listed under "Also reported by" with a count. The rating is the row's alone; no line carries one. An event whose sentence names a city or county belongs on that city's or county's row, created if the record lacks it, never on the state's row.

The rubric

Open. The jurisdiction lets data centers be built. It earns this rating from an approval of a project, rezoning, site plan or power contract; a tax exemption, abatement or payment in lieu of taxes; or a zoning rule permitting data centers by right or in a dedicated district.

Case by case. The jurisdiction decides projects one at a time without a settled rule. It earns this rating from a public hearing or comment period, a study or task force, a first reading or vote to draft rules or delay a vote, a bill passing one chamber, or a single site denial.

Restrictive. A rule in force narrows where or how data centers may be built. It earns this rating from a dated moratorium or temporary ban, or an ordinance, state law, executive order or utility tariff adding conditions such as setbacks, noise or water limits, or full power cost recovery.

Closed. The jurisdiction has shut the door on data centers. It earns this rating from a jurisdiction wide ban, a moratorium with no end date or one its source calls permanent or indefinite, or a denial with prejudice or after which the developer withdrew or abandoned the project.

A row is read from its own published events, each with its date and its source. Where they earn different words, the strongest act still in force sets the row, closed over restrictive over case by case over open, so a moratorium in force is not undone by a hearing or by one project's approval, and a moratorium that has ended no longer counts. Some events earn no word: a statement, public comment, a protest, a petition, a call on a body to act, a suit or a ruling in one, a developer's filing or open house, a bill that is not law and has passed no chamber, a bill vetoed, and an incentive cut or paused. On a state's own row only the state's acts count; a county's, a city's or another state's act is read on its own row. A moratorium whose source states no end date reads restrictive, and closed only where the source calls it permanent or indefinite; a denial whose source does not say whether the project can come back reads case by case. The rubric says what an event earns; the row's word is still a person's, given after reading the row's events and their sources.

Changes to this rule

• September 29, 2026 (later that day): one line per act, as the paragraph above says; the per event rating chips came off every line; and thirteen Texas events that name San Marcos, Taylor, El Paso, Bee County, Tarrant County, Westlake and Keller moved from the state's row to those rows, each logged on the corrections log.
• September 29, 2026: the rubric above written, one paragraph per word, and applied to the 108 ratings that had reached the scorecard as the classifier proposed them, before the acceptance rule of September 21. A rating the rubric supports from the row's own published events stays (44 rows); one it does not reads no published rating until an editor confirms a word (61 rows), each listed on the corrections log; one had since been confirmed by a person and two carried no rating. The four words had been defined in one sentence by the color each is drawn in, which read a moratorium as closed and new conditions as case by case; under the rubric a moratorium with an end date and an ordinance adding conditions both read restrictive.
• September 29, 2026: a row's rating no longer follows its events at all. Under the rule of September 8 the row took the rating of its latest published event that was not a planned date, so an approved event could move a row whether or not anyone had accepted its rating; now an event landing on a row never moves its rating, a row with no rating stays unrated, and a rating moves only when a person accepts one in the admin, which writes it onto the row after the approval.
• September 10, 2026: litigation became a timeline event category on both timeline entries and scorecard events, with the step recorded as filed, ruling, settlement or dismissed. A suit against an ordinance is recorded on the row and does not move its rating; only the jurisdiction's own acts do. Older rows keep their lawsuit label.
• September 8, 2026: the rule above replaced the earlier one, under which a row took the rating of its latest event of any kind, planned or not. The change followed a review that found nine rows reading from events that had not been approved; those events were returned to draft the same day and are listed on the corrections log.
• September 5, 2026: scorecard rows and their events first published, with the rating derived from the latest event.

Every project carries a cooling type: air, evaporative, closed loop liquid, immersion, or, where a source says the facility uses a liquid without naming which kind, liquid cooled with the fluid not disclosed. Not disclosed is the default and stays the default until a source states a method; nothing is inferred from a project's size, use or location.

Where a source names the fluid or the product itself, a glycol mix, an engineered fluid brand, that name is recorded separately as the cooling fluid, with its own source. A project can carry a cooling type with no named fluid; it never carries a fluid with no cooling type.

The cooling mix on a state, county or city page counts published projects by cooling type, the same way the use mix counts them by use, and the same rule applies: a not disclosed row stays on the table rather than being left off.

Changes to this section

• September 15, 2026: cooling type and cooling fluid added to the record.

A record page names, under Still not public, each of these it lacks and the kind of document that would carry it (October 6, 2026). The lines are the same on every record and live in one place in the code; a line never says the figure exists somewhere, only what the record will add the day a source gives it. Nothing is estimated to fill the gap.

Water use

A state water right filing or a water supply agreement.

Who would use it

A lease announcement or a tenant named in a filing.

Power the campus would draw

A utility or public service commission docket, an interconnection filing, or an air permit for an on site plant.

Who is behind it

A land record, a zoning application or a company's own announcement.

Jobs and tax

An incentive agreement, a county's economic development filing or a fiscal note.

The facts at the top of a record (who is building it, who is constructing it, how much power it needs and where the power comes from, the grid connection, how much land, how much water, the city, the county, the coordinates, the tenants) are each filled from one source at a time. A filing, a permit, an agency document or a company release can fill a field; press reporting on such a document can be cited on the timeline or as an outside figure, and cannot fill the field itself.

Every value carries the source that states it, linked from the fact, and a figure carries its basis: measured, what a permit allows, the developer's figure, or a third party estimate. A water figure keeps the unit the source states and, where the source says which, whether it is water withdrawn or water consumed. Coordinates carry their precision. The record's last verified date is the day a person approved the value.

A field is written by a person's approval, never by the classifier on its own, and each write keeps what it replaced, so a field can be returned to its earlier value. A field no source has filled reads Not disclosed; nothing is estimated to fill it.

The Record's own prose carries no hyphenated compound words: it reads "full time" and "940 acre", not the hyphenated forms, and the build stops on a published row that breaks the rule. A name is not a compound, so a hyphenated word that appears in a record's own name, in an operator or publisher name on that record's sources, or in a place name from the public gazetteers the Record already reads keeps its hyphens in the prose too. The list is checked before anything is split, and only whole words match, so a word that merely ends in a name is still rewritten.

Changes to this section

• September 18, 2026: a hyphenated word in a record's own name, in an operator or publisher name on its sources, or in a gazetteer place name keeps its hyphens. Two summaries that had been rewritten as "Mar a Lago" and "Bit O Wyo" read as their sources wrote them again; both changes are on the corrections log.
• September 15, 2026: contractor, site acreage, power source, grid connection and city added to the record; fields filled from a card's source with the source appended to the record's list.

A skeleton record, drafted from an operator's own list of sites or from the record's news monitoring, publishes once it carries every one of these: a name; a state; a county or a city; at least one source that is not a bare tier 1 query feed lead; and coordinates of any precision, a bare county centroid included. A record missing any of those stays a draft, worked on but not shown, until the missing piece is added. One rule in the record's own code is the definition, and the publishing step applies it to every draft after each batch of new records.

The rule is a floor, not a promise of depth: a skeleton that meets it shows what one source says and nothing more, and its page says it is a skeleton. Anything a source does not state reads Not disclosed. A record that has a county and no point is placed at the county seat and says so (below). Its status, statements and reference figures still go through a person before they publish; meeting the floor publishes the record itself, not what is said about it.

County level location

Where a source names the county but no site, the record's marker sits at the county seat, drawn with a dashed edge, and the map's facts panel reads "Location is the county seat; exact site not disclosed." The seat is a placeholder for the county, not a claim about the site; the seat coordinates come from Wikidata and the record names the seat in its location source. A town center reads "Location is the city center; exact site not disclosed." the same way. A parcel, address or permit point location replaces the placeholder as soon as a source gives one and an editor approves it.

What the sweep applies on its own

A draft skeleton that lacks only a source or only a place is searched for one. The search runs on the record's own monitoring worker, twice per record, and what it finds is read and classified the way every other item is. Two things, and only two, are then written without a person: the citation itself, meaning the article's publisher, title, date and link added to the record's source list; and a city the record does not name, where the source places the site in the record's own state. If the record then meets the floor above, it publishes.

Everything else the search turns up, a status change, something an official said, a figure, stays a pending item for a person to approve or dismiss, the same as any other item. The sweep never writes a status, a statement or a figure.

A share of these citations is drawn into the daily review sample. An editor reviewing the sample either agrees or overturns; an overturn removes the citation, returns the record to draft, and records what the rule got wrong. More overturns in a day than the limit stop the sweep from applying anything at all until a person turns it back on. The switch has an off position, and off is where it sits unless an editor has set it running.

Changes to this section

• September 29, 2026: a skeleton may be proposed from the EPA's facility registry, on a card an editor approves; approving it writes the record as a draft that meets this floor, with the registry's page as its source and the registry's point where the EPA placed it at the site. Locations are now described in six words (parcel, address, permit point, city, county, state), where the record had used city centroid, county centroid and county seat; the words changed, the locations did not.
• September 29, 2026, evening: a skeleton the registry proposes reads not yet verified, not reported, until a source states its stage, and it is listed under the count rather than in it. A registry card whose registry entry says data center by its NAICS code alone waits on a holding list for an editor's word before it reaches the review queue.
• September 17, 2026, evening: the sweep above, and the rule that it applies a citation and a city and nothing else, first written down and first run. Before this the same search ran one record at a time on a hosted runner and wrote nothing without a person.
• September 17, 2026, afternoon: a script places a draft that carries a city and no point at the city's own centroid (U.S. Census Bureau TIGERweb, the same source the operator census uses), and a draft whose county has no seat on file at the county's own centroid instead, both read as approximate; and the source requirement widened from tier 2 or tier 3 to any tier but 1, since a tier 4 local outlet is still reporting on the record rather than a bare search feed hit. The batch notes of that afternoon list what the wider rule published.
• September 17, 2026, morning: the rule above replaced the one written down the day before, which required a developer, a dated history entry and a tier 2 or operator source and had never been applied to a published record. This one has. The batch notes of that morning count what it published, state by state.
• September 16, 2026: the floor written down and county level placement added; proposed from a count of the draft records that day, not yet applied to any published record.

Every item the record's monitoring finds is a card, and every card is in one of three lanes. The third lane is always a person, card by card: a change of a record's stage to approved, running or cancelled; anything someone said; any reference figure; anything about one of the record's depth records; a possible duplicate; a scorecard event that would change a place's rating; and every question page. Nothing in it publishes, or is approved in a group, without an editor reading the card.

The second lane is everything else: cards approved in groups by kind and state, after the editor reads a sample of each group. Approving a group approves every card in it the way one card is approved, one build at the end, and a card the rules refuse as it stands stays out of the group until it is fixed.

The first lane is the four lowest risk kinds of card. A Daily Record news item, which is the article's headline, outlet, date and link with one sentence, publishes without a person from the start, because it only links out: every article the record's monitoring reads and finds to be about a data center gets one, as long as the source is a resolved outlet and not a search feed lead. The other three publish without a person once the record's own reading of that kind has agreed with the editor's decisions 98 times in 100 over at least ten of them: a source added to a row already on the record, word for word; a timeline line on a skeleton record from a primary source, the trade press or a local outlet; and a facility the EPA's registry lists, published as not yet verified. Every item added this way is listed on the corrections page the day it is added, marked added automatically. Ten a day are drawn at random for the editor to check, and an item the editor takes back stays on that list with the date. More than two of one kind taken back in a week puts that kind back in the second lane on its own.

Repeat reports fold before review, the way the scorecard and the statement lists fold them after it: when several articles report one statement, one figure or one act, one card carries it and every other article is recorded on that card as a corroborating source, so the record cites each and the editor reads the claim once.

Changes to this section

• September 30, 2026, evening: the Daily Record news item on every relevant article, published from the start (until then no article produced one, because the reading step never asked for it); the timeline lines on skeleton records taken from the trade press and local outlets as well as primary sources; repeat reports folded onto one card before review.
• September 30, 2026: the three lanes written down and applied to the review queue. On that day no kind had ten editor decisions behind it, so nothing yet publishes on its own; the first lane opens kind by kind as the decisions accrue.
• October 1, 2026: a news item passes the same checks as any line before it publishes (the text cleaner, no dash, every number in the sentence somewhere in the card's own facts), and a sentence that repeats the headline's or the article's words is held for an editor. Several reports of the same event on the same record within three days are one item, the others listed as also reported by. The first audit of twenty published items found none off topic, two repeating their source's words and two written from the headline alone; those kinds are held and written again.

How a bill is found. Once a day the record's own monitoring worker searches LegiScan for the same words in Congress and every state legislature: data center, data centre, large load, hyperscale, computing facility and server farm, and moratorium or sales tax exemption when either sits beside one of those. Only bills whose LegiScan change hash moved since the last look are read. A bill can also arrive by hand, from a roll call file read by a person. The words are kept in one place in the code so a term added there is searched everywhere at once.

What counts as a match. A bill matches when one of the core terms appears in its title or in the legislature's own digest of it. A paired term alone never matches. Whether a matched bill belongs on the record is always a person's decision, made on a card that asks that one question; nothing about a bill publishes before that. A bill kept off the record is not asked about again. Once a bill is on the record, a later stage from the legislature's own record may be approved under the sampled review bar; a bill signed, vetoed or taking effect is always a person's call.

What this search can miss, and how to tell the record. The search reads a bill's title and the legislature's own description of it, and nothing else. A measure whose data center provision sits only in the body of the text will not match: an omnibus act whose title lists a dozen subjects, and whose data center section is one of them, is exactly the shape the watch is blind to. The record does not claim to hold every such act. Where one is known to us it is read and filed by hand, and where it is not, the correction link at the foot of every page is the way to say so; a measure reported that way is read against its own text like any other.

A summary and a description are two different things. A description is the legislature's own line saying what the measure is about, and it sits under its own heading, credited as published by the legislature. A summary is an account of what the measure does, and it comes from four places only: the bill text, a Congressional Research Service summary, a legislative analyst's summary, or a fiscal note. It never comes from a press release or from coverage, and a bill whose summary names no such source does not build. Where the legislature's description is the bill's own title in the same words, the description is left empty rather than printing the title twice, and where no summary is on file the page says so and links the full text. The record's own sentence on each stage is written from the legislature's action record, whose wording is kept beside it. Hyphens in a description are split the way every line on the record is; the words are otherwise the source's.

What the measure is. Every measure says whether it is a bill, a simple resolution, a concurrent resolution, a joint resolution or a proposed constitutional amendment, and whether it changes the law: yes, no, or only if the voters approve it. That is read from the measure's own words and the rules of the legislature it is in, never from the letters in its number, because those letters mean different things in different places: Congress prints a House bill H.R. and a House simple resolution H. Res., and a joint resolution amends the constitution in one state and asks for a study in another. A measure that changes no law carries a line saying so, and the record never files it as a moratorium or as any other kind of law, whatever its words urge.

Subjects, and the record's own tags. The subject list on a bill page is the legislature's own, printed as it writes it. Beside it the record keeps a short list of its own tags, marked as the record's, for a thing a count needs to be able to separate. There is one so far: a bill about cryptocurrency or digital asset mining is tagged digital asset mining, so a mining bill can be counted apart from a data center bill. A mining bill is on the record where it regulates the facility as development or as a large electric load, or studies its grid or water impact; not where the subject is the currency, a tax on the activity, or what equipment may be bought.

How a vote is recorded. A roll call is read from the clerk's own file: the House Clerk's XML, the Senate's roll call XML, or LegiScan's roll call for a state legislature with the legislature's page as the citation where one exists. The question is kept exactly as the clerk words it, the requirement (a simple majority, two thirds, three fifths) is read from the clerk's vote type field and rendered as a plain label, and the totals by party are the clerk's. Every member's position is recorded exactly as recorded: Yea, Nay, Present, Not Voting. Where a clerk prints Aye and No, or LegiScan prints NV or Absent, the record stores the four words and keeps the clerk's own word beside it. There is no for or against column, no verdict, and no color that reads as a good or a bad vote. A member's name is as the clerk prints it. A position links to a speaker profile only where that person is already on the roster; no profile is created for a vote. A member's stated reason for a vote is a statement, recorded through the same path as every other statement, with the same treatment for every side.

Coverage. Each state's page carries one line: how many bills are tracked, whether the legislature is in session, when the record last checked and what it found. A check that finds nothing is recorded as a dated sentence with the words that were searched, so a zero never renders as a blank. Council and commission votes are not recorded here yet.

Credit. Bill data from LegiScan (opens in a new tab), licensed CC BY 4.0 (opens in a new tab). The record chooses the bills, writes each stage in its own words and cites the legislature's own page for every bill. LegiScan is credited as the finder on every such bill, every page and every export that shows one carries this credit, and the build refuses a page that shows one without it.

Changes to this section

• September 24, 2026: the credit to LegiScan made whole. It names LegiScan and its license, CC BY 4.0, with a link to each, says what the record changes, and now appears on every page that shows a bill, a vote or a coverage line LegiScan's data is behind, the state pages' Laws section and the Daily Record included, and in the header of every export whose rows it found.
• September 20, 2026: what the bill search can miss written down. The watch matches on a bill's title and the legislature's description of it, so an omnibus act whose data center section appears only in its text can be missed, and the section now says so and points a reader at the correction link.
• September 19, 2026: a description and a summary separated. Every bill on file carried the legislature's one line description in the summary slot, cited as a digest; it moved to its own slot with its own credit, and where it was the bill's own title in the same words it was left empty. The summary slot now takes four sources only and the legislature's digest is no longer one of them. Every measure gained a type and a plain statement of whether it changes the law at all, read from its own words; a resolution now says that it states a position and does not change the law. The record began tagging a mining bill so it can be counted apart from a data center bill. Twenty six federal measures were refiled under the numbers Congress prints, before any of them was published.
• September 18, 2026: bills and votes first recorded. The bill record, its append only history, the roll call and position tables, the daily watch and the coverage line written down here before the first bill was published; the first roll call on the record is House roll call 312 of September 16, 2026 on H.R. 9340, loaded from the Clerk's file with every member position.

A large number on the Record carries a plain comparison beneath it: a data center's capacity against New York City's summer peak demand, the electricity US data centers used against what an average home uses in a year. Every such comparison is arithmetic on two published rows and nothing else. One row is the figure; the other is a comparison unit, a fixed yardstick kept on the reference figures table with its own source and the page or table it was read from. The rule:

1. Capacity is compared only to capacity or peak demand. Megawatts and gigawatts against a city's or a state's peak load, never against energy over time.
2. Energy over time is compared only to energy over time. Kilowatt hours and terawatt hours a year against what homes use in a year, never against a peak.
3. Water is compared only to water, within the same scope. Withdrawal against withdrawal, consumptive use against consumptive use, a facility against a facility. A figure whose source did not say which measure it gave is not compared at all.
4. A conversion is arithmetic on two published rows, computed when the site is built and shown in the Details with both sources and the math. It is never a new fact and never becomes a row of its own.
5. When the two rows' years differ, the Details say so. The national electricity figure for 2023 is compared with the 2023 home figure, not the newest one.
6. A record's own figure converts only when it was measured or permitted. A developer's or a third party's estimate on a project record is never converted; the line under it says why. A national figure that is itself modeled may be converted, and its basis is shown.
7. The EPA's household yardsticks are the one exception to the water scope. The EPA gives what a family and a person use at home each day without saying how much of it is consumed and how much goes back, so rule 3 would compare nothing to them. A published national water figure that states its own measure is set against both, each with its year or the words "the EPA gives no year for it", and the same place says what each counts: the data center figure water consumed or withdrawn, the EPA's water used at home.

A figure stated in liters is shown in gallons as arithmetic: the liters divided by 3.785411784, the liters in a US gallon, and a year's gallons divided by 365 for a day, each step in the Details.

Mixing the kinds is how "six New York Cities" comes to mean three different things in three outlets. The Record uses one yardstick per kind and the same one everywhere. Another group's forecast is a reference figure of its own kind and renders in the forecast panel only, never in a count or a total.

The units on the record today:

• New York City (NYISO Zone J) annual electricity use, actual, 2024: 49,703 GWh a year (2024). NYISO, 2025 Load and Capacity Data Report (Gold Book), Table I-2, Baseline Annual Energy, Historical and Forecast, page 23, Zone J, 2024 row. The source (opens in a new tab).
• Water used at home each day by the average American family, more than: 300 gallons a day. EPA WaterSense, How We Use Water, the page's opening paragraph (page last updated August 4, 2026). The source (opens in a new tab).
• Average monthly electricity consumption per residential customer, United States: 863 kWh a month per residential customer (2024). EIA, Table 5A, 2024 edition, Table 5A, Residential Average Monthly Bill by Census Division and State, 2024 edition, U.S. Total row (page 2 of 2). The source (opens in a new tab).
• New York City (NYISO Zone J) summer peak demand, actual, 2024: 10,245 MW (2024). NYISO, 2025 Load and Capacity Data Report (Gold Book), Table I-4a, Baseline Summer Non-Coincident Peak Demand, Historical and Forecast, page 26, Zone J, 2024 row. The source (opens in a new tab).
• Water used at home each day per person, United States: 82 gallons a day per person (2015). EPA WaterSense, Statistics and Facts, Statistics and Facts, the first bullet under Water Use (page last updated March 11, 2026). The source (opens in a new tab).
• Average monthly electricity consumption per residential customer, United States, 2023: 855 kWh a month per residential customer (2023). EIA, Table 5A, 2023 edition, Table 5A, Residential Average Monthly Bill by Census Division and State, 2023 edition, U.S. Total row (page 2 of 2), read from the Internet Archive's capture of EIA's own PDF of May 24, 2025. The source (opens in a new tab).

The build refuses a page that shows a conversion whose two rows are of different kinds, whose unit is not published, whose water scopes differ or are not stated outside the household exception, whose figure is a project record's estimate, or whose arithmetic does not match the rows, liters to gallons included.

Changes to this section

• September 22, 2026: the rule written, with the comparison unit kind on the reference figures table and the build check.
• September 29, 2026: liters shown in gallons as arithmetic, and the EPA's household yardsticks set against the national water figure with the years of both and what each counts, on the editor's word.

A utility commission order, a tariff schedule, a rate class, a contract standard or a statute on how a data center or another large load customer is charged is a rate rule. It is recorded twice from the same document: as a law on the state's laws page, with what the text requires, and as a large load rule row for the electric bills question: the utility, the rule as the order names it, its status in the commission's own terms (proposed, approved, in effect, withdrawn, on appeal), the dates the text gives, what the data center pays directly and what stays shared with other customers, the minimum contract length, the minimum payment and the exit fee as the text states them, and the day the record last checked the docket. The rule row publishes with its law and never before; a field the text does not state reads not stated, and nothing is filled from coverage. A rate rule is a person's decision at every confidence: the sampled review policy refuses one on its own.

On that page a state reads not yet reviewed until the record has read its commission's orders or its statutes for this question. That is a fact about the record, not about the state: a state without a row may have a rule the record has not read, and the correction link at the foot of the page is the way to say so.

Changes to this section

• September 24, 2026: the rule written, with the large load rules table, the rate rule event type on the timeline, and the monitor's searches for a rate case, a tariff or a large load rule beside a data center.

An offer is what the companies behind a project have offered or agreed to give the people or the governments around it: a payment to each household, a community fund, money to a township or a school district, a credit on electric bills, a signed community benefits or host community agreement, a payment in lieu of taxes, or something in kind. Each is recorded as its own entry on the record's page, under What the developer has offered, from one source at a time: who made it, what kind it is, who receives it, the amounts, the condition, and where it stands (offered, agreed in writing, paid, withdrawn or lapsed).

Every number on an offer is one its source states. The amount to each recipient, the total, the amount a year, the number of years and the number of recipients are kept apart, and none is worked out from another: ten thousand dollars to each of about 4,500 households is not recorded as a total unless the source states the total. Where the source gives an amount only in words, the words are kept. The condition is kept in the source's own words beside it, because a developer's terms and the reporting about them can differ; a payment made when a building is finished is not the same promise as a payment made if a zoning change passes.

An offer is never added to another, never counted in a total on the site, and never counted as a tax payment, an investment or a figure about the project. A new status or a new amount is a new entry, with the earlier one kept and marked as replaced. An offer is published by a person's approval, never by the classifier on its own, and only on a published record.

Changes to this section

• September 28, 2026: the rule written, with the offers on the record page and in its JSON export.

A statement is what a named speaker said, as a source reported it, with the date and the source beside it. Statements groups every published one by who spoke: data center companies, government, chief executives, investors and finance, utilities and grid operators, candidates and campaigns, business groups and unions, advocates and residents, researchers and analysts, and other voices.

The group is read from the words the source gives for the speaker: the name, the title and the organization, as stored on the statement. The rules run in a fixed order and the first that matches decides. A newsroom or a writer is read first, then a candidate, a campaign or a party, then an office held now (a governor, a senator, a mayor, a commissioner, a judge, a public utility commission), then a resident or a member of the public, then a company named by name (its chief executive with the chief executives, its other people with the company, a utility's with the utilities), then research, public bodies, advocacy groups, business groups and unions, investors, and the generic words for a company last. A speaker who held an office before and is described by it now is read as a former official, in other voices. The data center companies include the developers named on the record's own projects.

A named person the source calls a company's chief executive in one statement is read as its chief executive in the statements that name only the company. Where the words name no group, a statement takes the group its speaker's other statements are read in, when at least two of them agree; otherwise the speaker type stored with the statement decides, and where that says nothing either, the statement is in other voices. A candidate who does not hold the office is not an official, and neither is a utility, a company, an analyst or a trade body (DECISIONS section 24).

The group is a reading aid, not a fact about anyone, and it is never a verdict on what was said. Where a speaker sits in the wrong group, the rule is wrong, and the correction link on the page is the way to say so. Topics on a statement are matched from the words of the statement itself, the same way as on the timeline, and are a finding aid only. The CSV and JSON of every statement carry the group and the rule that placed it.

A statement by a member of Congress about a federal bill or the country is filed on the national row and shown on the federal page; the speaker's own state page lists it under "What your members of Congress said", linking each line there, and no state's count carries it.

The same speaker saying the same thing on the same subject is one line on the record, dated to the earliest report, with every later report cited beneath it as "also reported by" and "said again on" its own date; a later report that takes a different position stands on its own and says what it changed from. Nothing is deleted, and a link to a folded report still lands on the line it joined.

Candidates on the November 3, 2026 ballot

Who is running comes from the official lists alone: each state's election office list for the November ballot, with the Federal Election Commission's candidate file beside it for the U.S. Senate and House, and each state's own list for governor. Office, district and party are as the list gives them, and every candidate page names the list and its date. A state legislative or local candidate is listed only where a data center source names them. A statement is linked to a candidate on name plus office or state, never on the name alone. The record looks for each candidate's public words on data centers in their own site, releases, newsletters, transcripts and video first, then interviews and news, then its own roll call positions; a candidate's own words, attributed exactly and dated, go on the record on their own, and a news paraphrase or an opponent's claim waits for a person. A campaign's claim about a vote is checked against the roll call as recorded. Every candidate in a race has the same space, in the order the list prints them, with no ranking, no stance word and no score; where nothing is found, the page says when the record looked and where.

The record follows each candidacy after the filing, through the primaries, withdrawals and the result. Every candidacy carries one status, the day it took effect and the source that states it: on the ballot, won the primary, lost the primary, withdrew, disqualified and, after the election, the reported winner, the certified winner or did not win. Two more words keep what the lists already said: did not qualify, for someone whose declaration the state rejected or who never made the printed ballot, which is never read as disqualified, and certified write in. A write in is shown on a race page, in a short group of its own, only where an official election list certifies the candidacy, with that list as the source: the state's own, or a county election board's where the state publishes none. A write in that no official list certifies is on no public page until a list names them, and a template's write in line is on none. A status is written only from an official source, and a county election board is one: its own list or results page is official for the races it administers. For a state or federal race the state's own list wins wherever the state publishes one, so a county board's list stands only where the state publishes none and is never cited over the state's for the same status. A candidate shown on a county list alone has a source line that names the board, as in "Franklin County Board of Elections, 2026 General Election Candidates", and where a county's title names no board the line names the county's election office. A news report or the Associated Press can show that something changed and is never the final word on an ending; the Associated Press or the state reports a winner, and only the state's own results page makes one certified. Nothing is guessed. A name a list leaves off is not a withdrawal, because a list that does not name someone does not say why, and where a source does not say whether a candidate lost the primary, withdrew or was disqualified, the last sourced status stands. Candidates who lost a primary, withdrew or were disqualified leave the race's cards and sit in a closed fold, "No longer in this race", with the status, the day and the source; what they said about data centers stays on their own pages, which ask search engines not to list them and stay out of the sitemap. After election day the winner heads the race page as the reported winner, with the source and the day, and then as the certified winner, with the state's results page as the source. The winner's sentence is in the page's title and description, and a reported winner the state does not certify is corrected on the corrections page. A page's last updated day is the day a status changed.

Changes to this section

• October 10, 2026: the candidate index for the November 3, 2026 ballot, from each state's official list.
• October 17, 2026: every candidacy carries a status with its day and its source, and the record follows primaries, withdrawals and, after November 3, the reported and the certified winner. It replaces the earlier line that the record did not follow primaries.
• October 20, 2026: a county election board counts as an official election authority, and its own list or results page is an official source for the races it administers. For a state or federal race the state's list wins where one exists, and a candidate shown on a county list alone has a source line that names the board.

Leaders: where a person stands on data centers

The Leaders pages show what an official in office or a candidate on the November 3, 2026 ballot has said about data centers and how they voted, from their own words and the roll calls, nothing judged beyond the labels the published rules below set. Who is in office comes from public domain and official sources: the current Congress and the national executive from the United States project's roster (public domain), the cabinet members who speak on energy and data centers from the White House's own page, each governor from the governor's own state site, and every candidate from the state's certified general election list, with the Federal Election Commission's file beside it. A candidate the FEC lists that no certified list confirms is never shown as on the ballot. Every office and candidacy row carries its source and the day it was verified. Portraits come from Wikimedia Commons under public domain, CC BY or CC BY SA, the license read from the file's own page; the record stores each one, never hotlinks it, and prints the author, the license and Wikimedia Commons under it. A campaign or office photograph is used only with written permission on file, never a news photograph and never a screenshot; a person with no such portrait gets their initials.

A statement gets one of four labels by the record's published rule, never by anyone's verdict. Two readings of the statement, each made on its own by a different language model (Anthropic's Claude Haiku and Claude Sonnet), propose a label. Where they agree, the label stands. Where they differ, a third reading (Claude Opus) reads the statement and the page it came from and decides, and the statement says the label was set by a tiebreak. The words that set the label are kept and shown as "Why this label". Every label sits beside its statement and its source, and any label can be challenged through Report a change on the page; a label a person changes is logged on the corrections page. Each label word on a page links to its definition here.

For data centers

The statement favors building, keeping, attracting or expanding data centers, or praises what they bring (jobs, investment, a tax base) without a reservation.

Raised concerns

The statement raises a concern about data centers (power bills, water, noise, land, secrecy, tax breaks, the pace), calls for a pause, a moratorium, limits, conditions, accountability or more study, or opposes one.

Mixed

The statement does both, with a real reservation on each side.

Not about data centers

The statement is not about data centers at all, even where the word appears. It stays on the person's page and is left out of the counts.

No label

A statement is held out of the labels, and still shows on the person's page unlabeled, only when its words are not clearly the speaker's own (an opponent's, a group's or the outlet's claim about them) or when its words, as the source reports them, support none of the labels.

The only sentence that summarizes a stance is one template, the same for everyone, built from the counts and the labels alone. Once a person has three or more labeled statements it reads, for example, "Of 12 statements on the record, 9 raised concerns about data centers and 3 were mixed." A statement repeated across outlets counts once. Below three, a race page shows the candidate's newest statement in their own words, with its date and outlet; with none, it says when the record last looked. A measure carries one of three effects, helped data centers, limited data centers or neither, set by the same two readings and the same tiebreak, and only for a measure with a final passage roll call or a signature. A vote counts toward what a person did only on a final passage roll call; a cloture motion or an amendment is listed and not counted. A vote row reads plainly ("Voted no on a measure that helped data centers"); a yes on a measure that helped data centers, or a no on one that limited them, counts as a vote in favor of data centers, and the reverse as a vote against. Nothing is scored, ranked or sorted by stance; party is a word and never a color; every candidate in a race has the same card in the order the official list prints, or alphabetically where it prints none. Every race page and person page says the day the record last checked it. The "Where we looked" panel on a person's page lists each source the record swept and the day it was checked, the sources that held nothing included. A person's page with no statement and no vote is kept out of search results until the first one lands, and stays linked from its race page. The hub's map shades a state by how many of its ballot candidates have a statement on the record, never by stance. The hub's search reads a static index of every person, race and state on these pages.

Every race lists its candidates in one order: the candidate in office first, then the Republican and Democratic nominees together by last name, never one party ahead of the other, then everyone else by last name. Inside each of those groups a candidate with a statement or a vote on the record comes ahead of one with nothing. No poll, forecast or rating affects the order.

On a state's leaders page each candidate in a race carries one word, read from the same counts as the sentence and nothing else. For, Concerns or Mixed is the label most of the person's labeled statements carry, and a tie for the most reads Mixed. Few on record, with the number of labeled statements beside it, means fewer than three carry a label, too few to read. No statement means the record holds no statement from the person on data centers. The word is a count, not a verdict: every candidate's word is printed the same way, with no color, and the race's line counts how many candidates read each word.

Changes to this section

• September 25, 2026: the rule written, with the statements pages and the speaker type the review console writes from the same reading.
• September 26, 2026: the candidates group's page lists its statements by the state of the race, the office sought and the candidate, read from the words the source gives for the speaker; a statement whose words name no office or no state takes the one race the same candidate's other statements name, and otherwise sits under office not stated. The race is a reading aid in the same way as the group.
• September 29, 2026: the front page's What was said shows no two statements from the same speaker or the same campaign, and where the record holds statements from more than one side on the same matter, one from each; the vote check below.
• October 10, 2026: a repeated statement is one line on the record itself, said again on its later dates; the candidate index for the November 3, 2026 ballot.
• October 12, 2026: the Leaders pages, the people roster, the stance labels and the measure effects a person accepts, and the templates above.
• October 13, 2026 (evening): a state's leaders page is an index of its races, each candidate with the one word above; the statements themselves are on the race and person pages.
• October 13, 2026: labels and effects set by the published rule (two readings, a tiebreak that reads the source) where a person had accepted each; one summary sentence for everyone in place of the earlier templates; portraits from Wikimedia Commons with the author and the license under each; only a final passage roll call counts toward what a person did.
• September 29, 2026 (later that day): one statement per speaker per claim on every list. Where the same speaker made the same claim and several outlets carried it within thirty days, the lists show it once, with the count of sources and each one linked under "Also reported by"; the carrying source is the speaker's own page or another primary source where one exists, else a quoted statement over a paraphrase, else the earliest. Every stored row keeps its own anchor inside the carrying card. A statement's state is the state of the office its speaker holds, read from the title and organization the source gives ("U.S. Representative (D-Arizona)", "U.S. Senator from Ohio"), so a member of Congress quoted in another state's story appears on their own state's pages; where the words name no state, the statement stays in the state it was reported in, which is where a campaign spokesperson's claim about an opponent stays, attributed to the spokesperson.

What was said on the front page

The front page shows up to four statements, newest first, from sources the record resolved and about data centers; where fewer than two qualify, it shows the newest of the last 30 days. No two come from the same speaker or the same campaign: a candidate's campaign staff and committee speak for the candidate, and a statement signed by several speakers counts for each of them. Where the record holds statements from more than one side on the same matter, the page shows one from each, together, under the matter's name; one side of a matter is never shown alone while the record holds another. The matter is what the record stores beside the statement: the race a candidate, a campaign or the person a campaign speaks for is in, then the record the statement names, its bill, or the scorecard row it is filed under. In a race each campaign is a side; on a record, a bill or a place a side is who is speaking, by the groups above, with a company, its chief executive and its investors one side. It is a reading of who spoke, never of what they said.

A statement about how a lawmaker voted

Where a statement says how a named legislator voted on a named bill, the record sets its own roll call beside it, in one of three words. Confirmed by the record: a published roll call on that bill lists the member voting the way the statement says. Contradicted by the record: the record's roll calls on the bill list the member, and never that way, whether the other way, present or not voting. Not on the record: the record holds no such bill, no roll call on it, or none that lists the member. The bill and the roll call are linked.

The claim is read from the statement's own words by fixed rules: a person named before "voted", with a title, a description between commas or a parenthesis set aside, and he or she read as the speaker; the way they voted (for, yes, in favor of, in support of, to advance or to pass; against, no, to reject); and the bill after it by its number, its title or both, with the year where the words give one. A statement that says a person would have voted, will vote, sponsored or voted, or voted for a bill it does not name, claims nothing that can be checked and gets no line. Where the words give a number the record does not hold and a title it does, the title decides and the line under it says so. The member is matched on the clerk's printed name and the state, and where two members fit and nothing tells them apart, no line is shown. The check says only what the roll call the record holds lists; it is never a finding about the speaker or the statement beyond that.

A located record shows the distance from the site to the nearest homes. The figure is computed, not estimated, and this is how.

The unit is the 2020 Census block, the smallest area for which the U.S. Census Bureau publishes a count of housing units. Every block within five kilometers of the site with at least one housing unit in the 2020 Census is retrieved from the Census Bureau's TIGERweb service. For each, the shortest great circle distance from the site's coordinates to the block's boundary is computed, zero if the site lies inside the block. The smallest of those distances is the figure shown, with the block's housing unit count beside it.

The site is the record's coordinates. For a parcel level record that is the parcel's center, so the figure is a distance from the middle of the parcel rather than from its nearest edge, and the parcel outline on the map shows the difference. For an address level record it is the address point. No figure is computed for a record placed at a city or county centroid, because a distance from an approximate point would be an estimate.

A block with housing units is where the Census counted homes in April 2020; it says nothing about homes built since, and a block boundary is not a front door. The figure is a public domain measure of how close the nearest counted housing is, and each record links the query that produced it.

A five digit ZIP code typed into a search box is looked up in the U.S. Census Bureau's ZIP Code Tabulation Areas, public domain files: the 2020 Gazetteer gives each area's internal point, and the Bureau's 2020 relationship file between those areas and counties gives the county that holds most of the area's land. The result names that county, opens its page where the Record has one and the state's page where it does not, and lists the nearest data centers on the record by straight line distance from the area's point, within 150 miles.

A tabulation area is the Bureau's approximation of a ZIP code's area, not the Postal Service's, so a ZIP that covers only post office boxes or a single organization has no area and the search says so. An area that crosses a county line is filed under the county with most of its land. The Record names no city for a ZIP, because the Bureau's files do not tie one to an area reliably. A distance is measured to a data center's site where the record has one, and to the center of its town or county, named in the result, where it does not. The position of a ZIP is never sent anywhere: the lookup reads two small files from this site.

What the site collects about its readers, why, and which companies handle it is on the privacy page, which describes only what the code does. Every change to that page is listed here with its date.

Changes to the privacy page

• October 20, 2026: a choice about visit counts. Session recording and heatmaps are not used, and two analytics services count visits, Google Analytics and Ahrefs Web Analytics. In the European Economic Area, the United Kingdom, Switzerland and Quebec, and wherever the site cannot tell where a visitor is, neither loads until the visitor accepts; elsewhere both load and can be turned off. Global Privacy Control is honored. The Privacy choices link at the foot of every page changes the choice, which is kept in the browser under dcr-consent and is not a cookie. Search terms and lookups are kept out of the analytics. A section for readers outside the United States added. The line saying a consent prompt was coming removed.
• October 16, 2026: local alerts added (an email when a data center is proposed, opens, or closes in a place or within 25 miles of it), with what they store and how long, and the newsletter's weekly track named the Weekly Record.
• October 13, 2026 (night): the line promising quarterly figures on contributions removed; the site publishes no figure about contributions.
• October 13, 2026 (evening): Cloudflare Web Analytics named, which sets no cookies and collects no personal data.
• October 13, 2026: the publisher's mailing address added (Data Center Record, PO Box 1564, Corvallis, OR 97333), which every email now carries in its footer, and the line on the forms' Turnstile key reads the key the site runs on.
• September 17, 2026: the privacy page first published. It names the three analytics scripts, Cloudflare hosting and Turnstile, the forms and their 12 month retention, the newsletter and its 30 day deletion after an unsubscribe, and Stripe contributions.

Errors are corrected openly. A correction is added to the record it affects, with the date and a note describing what changed, and it is listed on the corrections log. The earlier value is not quietly replaced.

Because records are kept in version control, every change to the dataset carries a timestamp and a visible difference. The correction history is not something the Record maintains voluntarily; it is a property of how the data is stored.

To report an error, write to corrections@datacenterrecord.org with the record and the source that shows the correct information.