WhatsTheHost Research
dry-runupdated Sep 5, 2026

Methodology

How the numbers are made, and how to check them.

Written for an editor who has five minutes and a reporter who has fifty. The short version: every number is either computed by published code from federal data, or traced to the exact sentence of a source you can open in one click. The long version follows.

Pipeline

The five gates

1

Capture

Every source is registered with an ID (S001, S002 …) and a tier. Tier 1 is a primary document: a federal dataset, a system-operator filing, a regulator's order, a company's own release or filing. Tier 2 is a national laboratory, standards body, peer-reviewed or preprint paper, or a transcript. Tier 3 is trade press, which may support a claim but must be replaced by a Tier 1 or 2 source before the claim can appear in a headline, a chart or a pitch.

2

Evidence

Every cited sentence is re-fetched from its canonical URL and machine-checked against the live page text on a stated date. VERIFIED_LIVE means the exact string was found. PARTIAL means the substance is supported but the wording, unit, date or framing differs, and the study's text was rewritten to match the source; the discrepancy is recorded. NOT_FOUND, PDF_OR_PAYWALL and FETCH_FAILED mean a human must verify before publication, and such a claim may not headline. Every claim carries a deep link that opens the source at the cited sentence.

3

Claim ledger

Every factual claim in a story gets an ID (C-001 …), its source, its verbatim (character for character, or null), its status, any discrepancy between what we first wrote and what the source says, and the standing instruction for anyone editing the story. The ledger is published with the story.

4

Inference ledger

Every interpretation, prediction and piece of speculation gets an ID (I-01 …), a confidence grade (HIGH, MEDIUM, LOW, SPECULATIVE), the claim IDs it rests on, the named reasoning model it borrows and whose it is, its reasoning, a dated and observable falsifier, and the strongest person or argument that disagrees. An inference with no falsifier does not ship.

5

Adversarial audit

Two reviewers with opposite briefs read the finished story: one on the data (duplicates in a percentile population, misaligned windows, numbers that cannot be reproduced from the stated inputs, charts whose visual claim exceeds the data), one on opinion and attribution (positions attributed without a document, confidence not supported by evidence, falsifiers that are not observable, models misapplied). They report defects only, graded BLOCKER, MAJOR and MINOR. A story with an open BLOCKER is not published. The findings and the fixes are published in the story's audit history.

Rules

The eleven standing rules

These apply to every desk, every story, every morning. Breaking one is a defect and goes on the corrections page.

  1. No claim without a ledger ID.
  2. Verbatim is character-for-character or null. Paraphrase is never inside quotation marks.
  3. Unverified claims never headline, never chart, never pitch.
  4. No unverified claim inside a HIGH-confidence inference.
  5. Every inference has a dated, observable falsifier.
  6. Positions are attributed to named living people only with a document and a deep link.
  7. Derived metrics are labelled WhatsTheHost computations, with window and inputs stated.
  8. Percentile populations contain no duplicates.
  9. Audit changes are published in the story's audit history.
  10. Nothing goes to production and nothing is sent without the publisher's explicit go, per action.
  11. No invented first-person experience. Unconfirmed first-person material is removed before publication.

The Power Index

Sources and formulas

Grid layer

U.S. EIA, Hourly Electric Grid Monitor (EIA-930 BALANCE files), EIA-930. Window: Jan 1 00:00 UTC through 08-18 04:00 UTC, identical hour count every year, Feb 29 excluded.

Price layer

EIA Electric Power Monthly Table 5.6.A, industrial sector, May 2026 (released 2026-07-23, preliminary, Form EIA-861M), table_5_06_a.xlsx. All prices are nominal; no inflation adjustment is applied anywhere.

Carbon layer

EPA eGRID2023 Rev.2 Summary Tables, Table 1 (released 2025-06-12), summary tables, blended with the observed 30-day generation mix from EIA-930 times EPA fuel-specific CO2 rates (coal 2,270, natural gas 900, petroleum 1,670, other 1,200 lb/MWh).

Scoring
Percentile rank 0-100 across the UNIQUE (state, balancing authority, eGRID subregion) units in this index, then joined back onto every metro sharing that unit. 100 = most favourable to a buyer adding load. Metros that share a state and a grid share a score by construction; the metros_sharing_this_unit field says how many.
Composite
AI Power Readiness = 0.45*Cost + 0.35*Grid Stress + 0.20*Carbon. Weights are an editorial judgement, disclosed and adjustable in the published tool.
Grid stress
Grid Stress = 0.50*seasonal peak headroom + 0.30*inverse ramp volatility + 0.20*inverse day-ahead forecast error. Balancing-authority level, not metro level. Headroom compares the trailing 30-day mean of daily peaks against the 99.9th-percentile hour observed in the SAME calendar window across prior years, so a winter-peaking grid is not credited with headroom for being measured in August. Ramp volatility is computed only over consecutive-hour pairs. Forecast error above 15% is floored at the worst rank. The 365-day peak is the 99.9th percentile hour, not the maximum, to exclude single-hour data errors. Balancing authorities more than three days behind the freshest one are dropped before the aligned cutoff is chosen.
Carbon score
Carbon score ranks the OBSERVED intensity computed from each balancing authority's measured 30-day generation mix times EPA fuel-specific CO2 rates. The EPA eGRID subregion figure is published alongside as the citable reference. Where the two diverge by more than 25% the row carries a carbon_flag.
Cost scenarios
Energy-only cost. PUE 1.3, 100% utilisation, 730 h/month, state industrial retail price. Excludes demand charges, capacity charges, transmission riders, negotiated contracts, taxes and incentives. Formula: MW × 1,000 × PUE × 730 h × 12 months × state industrial rate.
Local constraints
The Local Interconnection Reality column is qualitative and sourced but deliberately excluded from the numeric score. Markets sharing a balancing authority and a state share identical grid and price inputs; that column is where they diverge.

Known limits

  1. Observed carbon intensity is computed from generation INSIDE each balancing authority. It ignores imports and exports, so an import-heavy grid understates the carbon content of the power it actually consumes. The eGRID subregion figure is published alongside for that reason.
  2. Scores are computed over unique state-plus-grid units, not over the metro list, so the ranking cannot be moved by adding or removing metros that share a grid. Metros sharing a unit share a score, which is a real limitation of resolution, not a bug.
  3. There is no grid-slack component in the stress score. Flat load growth is indistinguishable in this data from saturated load growth, and treating slow growth as available headroom scored a state under a construction moratorium as the least-stressed grid in the index.
  4. State-average industrial retail price is a proxy. No large data center pays it. Real deals are negotiated, often with special contracts, and increasingly with dedicated large-load tariffs.
  5. Balancing-authority grid conditions are broader than a metro. PJM covers 13 states; Ashburn's local constraint is a substation and transmission problem PJM-wide data cannot see.
  6. eGRID2023 is the current vintage as of Aug 2026, so carbon intensity lags the grid by roughly two years and understates recent solar and storage additions.
  7. Metros spanning multiple balancing authorities are mapped to the primary BA carrying the data center corridor; the alternate is disclosed per row.
  8. Forecast error above 15% is flagged LOW_CONFIDENCE and excluded from the stress score, because it usually indicates sparse day-ahead forecast publication rather than genuine operational surprise.

Code and data

The computation scripts (process_grid.py for the grid layer, markets.py for scoring and the CSV) are published with the repository for this site. The dataset is released as CSV and JSON under CC BY 4.0 from each story page. Anyone should be able to re-derive the headline number from the CSV in five minutes; if you cannot, that is a defect and we want to hear about it.

What AI does and does not do here

AI agents run every stage above: they fetch, compute, verify strings, draft, and audit one another. They do not decide what is published; a human publisher does, after reading the ledgers. They do not supply numbers; numbers come from data and code. They do not supply quotes; quotes come from documents and are checked against them. Where AI drafting introduced an error that survived to publication, it is recorded on the corrections page like any other error.

Reviewers

The methodology is open for external review. Reviewers read this page and one complete story with both ledgers, and their names and review dates are published here. As of Sep 5, 2026: no external reviewer has yet signed. We are seeking two to three (utility rates and tariffs; grid operations; data center development). Names and dates will appear here when they do.