# Process memo — Data Centers and Retail Electricity Prices (2026-10-01)

## What was done, in order
1. Setup (2026-10-01 19:45 EDT, unattended mode-1 web commission): request recorded verbatim; Setup Brief; Argument Sketch; repo scaffolded (nber, LaTeX-first); auditor APPROVED.
2. Literature: 56 PDFs acquired by the dossier toolkit with converted text; 7 paywalled/unavailable logged (Scalera et al. CEPR DP21854 abstract only; Muller NBER 35100 and Knittel et al. NBER 34065 403 to scripts); Literature Map v2; auditor REVISE on three stale lines, fixed.
3. Frontier data (Stage 1D): data/SCOUT.md; EIA-861 1990–2024 parsed by the paper's own loader and validated against macrolib.eia861 on 2012–2024 to 1e-14; EPA ICIS-Air + FRS facility list (623 facilities, 182 counties) with the 2014-migration censoring rule and the known-cases test (15/31; dates 1.5 years late, never early); CBP 1986–2023; ATLAS map as a second universe (IM3 not obtainable by script). No scraper was needed.
4. Model (Stage 1T): model/model.py, 18 claims green in derive.py; display equations tied to checked expressions by equality claims.
5. Identification Memo committed before any result (commit f696085).
6. Estimates (Stage 1E): blocks A–J of code/03_analyze.py; the engine's econ shelf (one engine fix: csdid clustering on a non-unit column with `differences` 0.3.0).
7. Plan v2 → outlines 3A/3B/3C → blind referee memo on 3C → the Plan-gate loop: the memo applied the memo's own power rule and found the null uninformative at the median dose; four design additions (dose-response on the stack with a cross-checked main effect; three-level facility class; drop-future-cohorts; utility-level regime split) and two measurement corrections (modal ownership per county; years 6–8 beside 0–5).
8. Draft v1.0 as tokens (no typed numbers), 14,000+ main-text words, all NBER floors met.
9. Attempt 3 (resumed after a usage-limit interruption): block J of 03_analyze.py had failed on a key error and the dose-response had never run; once run, the residential slope in the load-share proxy is flat (−0.02 log points per percentage point, s.e. 0.08) and only the industrial slope is negative, so the "gradient" claim was rewritten to the facility side (Clean Air Act class, third facility, the municipal cell) and the proxy's limitation (it varies only with the serving utility's size) stated. Every typed data-construction count became a token on a new result key. Auditors: Draft v1.0 APPROVE; 3C re-check REVISE on the outline record (fixed); Data REVISE (final clean run owed — done before delivery); Model REVISE (the general-allocator equation and the demand-response term were hand-typed in Appendix A) → both are now checked claims with display fragments (derive.py 28/28); Lit Review re-check APPROVE. Quality passes v1.1 and v1.2; the full text gates (passport, typography, reverse sweep with the PDF, quotecheck, paperstats, declaude, bibcheck) clear.
10. Referee round 1 (five blind + case-for) → the loop-back (Round 1 dispositions.md) → Draft v1.3; referee round 2 (five fresh + case-for) → Round 2 dispositions.md → Draft v1.4; the second citation audit (a diff); the fresh-context claim check (every registered number re-read from the register; the known-case in-service years corrected; the first-stage sentence corrected); the copy-critic pre-send pass; a final clean run of run_all.sh as the replication evidence; the Submission build; the replication package; the readings filed to the Library.
12. Attempt 4 (resumed 2026-10-02 after a second usage-limit interruption): the number-auditor (CLEAN: 383/383 claims, 0 orphans in the reverse sweep, 22 hand spot-checks) and the Final auditor (REVISE on five record and packaging items, none in the paper's text: an exhibit-to-script map added to the replication README, the number-audit record saved, checklist rows brought into step with the files, re-run gate outputs committed, a stray bytecode file removed from the LaTeX bundle; the last two fixed in the engine so they cannot recur). Every delivery gate re-run clean before the result was written.
11. Engine fixes made on the way, committed to paperwright: the reverse sweep's PDF pass treats a subsection heading's number as structure; the generated preamble tells siunitx to ignore input commas (a count of 2,445 had printed as 2.445) and numbers appendix exhibits A1, B1, …; the econ shelf's csdid clusters on a non-unit column (attempt 2).

## Decisions and why
- County design is the headline (pre-specified); the utility-unit version is a reconciliation table, because its treated/never-treated pools differ in size by two orders of magnitude and its industrial series fails the pre-test.
- Texas dropped (territory coverage 34 percent); always-treated counties dropped; 2025+ cohorts kept as not-yet-treated.
- Virginia synthetic control reported at 2015 as primary; 2008 confounded by Dominion's 2007 re-regulation.
- The honest (Rambachan–Roth) bound is reported beside the naive bound wherever a dollar figure appears.

## What the referees converged on
Round 1 (five blind referees under sampled dispositions, all "major revision"; acceptance odds as-is median ~30 percent, ~60 percent after revisions; the case-for reader voted accept conditional on light edits): (1) the dose was assumed at 100 MW while the typical first facility is small, so the premium bound and the "paid its way" sentence were not earned and there was no first stage; (2) the utility-unit residential estimate in vertically integrated states (+2.8 log points) sat in an appendix table undiscussed; (3) the heterogeneity cells carried state-clustered stars on 4 to 34 treated counties; (4) the informative-null comparison used long-run magnitudes against a six-year window; (5) the commercial class vanished after the headline table; (6) the composition reading was asserted rather than quantified; (7) θ = 1 with κ = 0.5 contradicted the paper's own class prices. The loop-back did the work for every one of these (Round 1 dispositions.md; Identification Memo amendments); the abstract, introduction and conclusion were rewritten to the bound the design supports. After the bootstrap inference was added, the exploratory cells the earlier draft leaned on (major-source first facility, municipal systems) no longer distinguish from zero, and the paper says so.
Round 2 (five fresh referees, all "major revision"; odds as-is median ~30 percent, ~60 percent after revisions; the case-for reader again accept conditional on light edits): the first stage the round-1 loop added became the center of the critique — its sentence was arithmetically wrong, and read correctly it rejects a 100 MW first facility, which is the load the inversion assumed; the county-to-utility sequence should be shown in one place and the utility unit treated as the estimand; M̄ = 0.5 is a choice; the raw-paths figure disagreed with the event study; counts in the data section had template residue; θ_I conflates cost causation with allocation; the rate-case hazard is asymmetric in practice. Dispositions in Round 2 dispositions.md: the first stage corrected and given exhibits with the load it permits; a sequence table and a utility-first-only specification; M̄ = 1 in the abstract and the set on the shorter window; the raw figure redrawn as a raw difference-in-differences; counts regenerated; the caveats on θ_I and the lag stated. Not done, and open: a headline at the utility unit (six comparable never-treated utilities); rate-case filing data; the Scalera et al. full text; Bonbright/Kahn/NARUC citations from a library copy; a measured facility load.

## Unresolved issues and limitations
- Table 2 reports 60,754 industrial county-years (the panel) and Table 4 60,750 (the estimation sample after the estimator drops four); both trace to code and the prose cites neither.
- After two referee rounds the open items are: (1) a headline at the utility unit — the model's unit — which the data cannot support because the like-for-like never-treated pool is six utilities; the paper shows the full county-to-utility sequence and leaves the vertically integrated residential sign open; (2) the facility load, which no public record carries; the first stage bounds the typical first facility at roughly 20 MW and the premium is bounded only on an assumed load; (3) the asymmetric rate-case hazard, stated as a limitation without filing data; (4) the Scalera et al. (2026) discussion paper, read only in abstract (CEPR serves the PDF behind a login); (5) Bonbright, Kahn and the NARUC manual, which a referee asked to be cited and which are not on disk or at a URL the engine could fetch — a co-author with a library copy should add them; (6) the appendix derivative fragments print in the derivation checker's algebraic ordering.
- The exploratory cells the earlier drafts leaned on (major-source first facility, municipal systems) do not survive wild-cluster bootstrap inference on their handful of treated states; the paper reports them as signs.
- The facility list's coverage varies by state (Oregon, Utah, Illinois, New York misses); dates run late.
- The utility-level industrial rise is not reproduced by the county design; its likeliest cause (Part C mix in restructured states) is stated, not proven.
- No facility MW: the dose is a proxy (100 MW ÷ serving utility's sales; the first facility's Clean Air Act class).
- The 2025–2028 wave has no post period.

## Where more or better data would change the answer
- Utility-level load by customer and the billing class of each facility; facility MW from state permit portals (Virginia DEQ, Maricopa, BAAQMD); the IM3 atlas; interconnection-queue dates for the post-2021 wave.

## Construction cost (Stage 1D)
- Scout and build: roughly five hours of agent time across two attempts (the whole paper took four attempts and about ten hours of agent time, two referee rounds included); the simpler version would have used macrolib's 2012–2024 county prices and a hand-typed list of ~30 hyperscale openings — a 12-year panel and 30 events instead of 24 years and 148.

## Asks a co-author could fulfill
- A Virginia DEQ permit pull for generator kW in Loudoun, Prince William and Henrico; a FERC Form 1 match for the serving utilities' rate-base additions by year.