Origin
The issue began with a promise nobody could check. It ended as a three-rung ladder: a rule, a receipt, and a say.
The question was chosen before any conclusion existed: when AI data centers place large new demands on electricity grids, how should policy decide who pays? The occasion was concrete — household bills rising across the largest American grid region while seven AI companies signed a public pledge that households would not carry their costs.
Four rival endings were then built and attacked as a slate. One — a "watch the docket, distrust both panics" hedge — was killed for surviving only by asserting almost nothing. Three survived wounded, and the author selected a fusion of two: an enforceable-rule standard (a promise is not a rate) joined to a citizen-legibility standard (filed in a docket is not visible to a household).
Two outside critiques, run by the author in a frontier model and adjudicated finding-by-finding, then changed real substance: they caught a category error in the essay's headline dollar figures, and they falsified an early "nobody publishes this" claim with two industry-funded studies — which the published essay now carries as the companies' strongest evidence.
Where it came from
Human trigger
A March 2026 pledge from seven AI companies promised to protect household ratepayers. The author's question: which parts of that promise have become rules a company can be held to — and how would anyone outside the deal know?
Process pressure
The landing was selected from rival candidates spanning both directions — including a pro-build case the essay must carry at full strength — and the author's own compression rounds repeatedly cut structure that read as machinery.
Public check
The claims were settled by direct reads of primary documents: Virginia's GS-5 rate order, the JLARC study and its projections table, the market monitor's auction analysis, both industry-funded cost studies, and the Utah statute.
The conversation behind this
How did the issue change?
Read the extended public-safe development record: the author-AI exchange across the issue's working sessions — the slate selection, the manual-critique correction that reshaped the process itself, the caught errors, and the repairs, in chronological order.
What shaped it
Changed
- An outside critique caught the essay presenting one auction year's estimate and a three-year cumulative estimate as a single range — a category error. The published essay never merges them again.
- An early claim that nobody publishes cost-versus-payment accounting was falsified by two industry-funded studies; the essay narrowed its claim to what is actually missing — recurring, regulator-required, after-the-fact public accounting — and adopted the studies as the companies' best evidence.
- The final integrity review caught a temporal impossibility — an "audited ledger" appearing before any decision it could audit — and the essay split its central instrument in two: a price tag before the decision, a receipt after service begins.
- Mid-issue, the author corrected the process itself: external critiques must be run manually by the author in an outside frontier model, never executed by the in-house pipeline. Two gates were re-run under the corrected rule, and the re-runs became the reviews of record.
Prevented
- A landing that implied the pushback is winning or the buildout is slowing — the record supports neither.
- Treating "no one can verify the costs" as proof that households are currently subsidizing data centers. Unverifiability is the problem named; it is not a verdict.
- A one-sided reading of Virginia's own study — its "data centers are paying their full cost of service" finding travels everywhere with its modeled bill projections, both halves in view.
- Process artifacts being used as proof that the essay is true.
Preserved
- The strongest case for building, stated at full strength before the essay asks who pays.
- The author's protected emotional beats — including the household that never asked for a better explanation, only a bill that does not rise for someone else's project.
- The dividing line: can someone hold the company to it if the company changes its mind?
- Five published failure conditions that would prove the essay wrong — including the one where Virginia's new framework produces the missing receipt and the essay fails "in the best way."
Paired opinion essay
After publication, the author took the same verified fact base into the lighter-rigor sibling letter: The Black Box Next Door (Intuition, July 2026) argues the one-sided bet this audited essay does not make — that secrecy, not cost, is the main driver of local opposition, and that full transparency would raise acceptance. The pairing shows the same facts carried at two levels of rigor: the audit here, the opinion there. The Intuition page carries its own origin note, including the multi-model drafting experiment behind it.
What this page is not
This Origin page is context, not certification. It shows what shaped the issue and where the process applied pressure. It does not prove the essay true, safe, careful, or complete.
The written issue remains canonical. Use the Structure, Reference, and Crosswalk layers to inspect how claims and sources were separated. The author's sealed pre-process note is an internal diagnostic and is not published; its existence and its match behavior are disclosed — and explicitly not treated as validation — in the probe-match methods note and the development record.