The changes tie to the hundred. The baselines move.

Georgia 55378 · NCUC E-100 Sub 208A · 2026-08-04

Nine filings: five in Georgia’s quarterly large-load docket, four in the Carolinas’ semi-annual one. Of the twelve changes these filers state from one period to the next, ten reconcile exactly and two are out by 100 MW — one rounding step on a table headed approximate. What moves without being stated is the baseline they are measured from.

Georgia Power’s Q3 2025 large-load report reached the Commission with 535 change flags in the public version of its attachment. That attachment has 107 project rows and five change columns. 107 × 5 = 535. Every cell said Y. The Q4 report, filed 11 February 2026, carried 530 flags across 106 rows. Same arithmetic.

The company found it and said so.

I compared the corrected attachments against the originals as multisets of numeric tokens, page furniture stripped: 3,078 numbers in Q3 and 3,561 in Q4, the same numbers the same number of times before and after. A multiset would not catch two figures swapping rows, but nothing entered either attachment and nothing left it. What was wrong was the report’s own answer to what changed since last quarter, and it said every project had changed everything. Q3 stood that way for 103 days, Q4 for 21.

How this was done

Each PDF came from the commission’s own document handler, converted with pdfminer.six and saved as the extractor emitted it, ligature damage and all. Three operations produce every number here. Flags: count the lines whose characters, whitespace removed, are exactly Y. Numbers: drop the page furniture — the watermark, the PUBLIC DISCLOSURE stamp, and the running header, which prints on its own line sometimes as Page n of m, sometimes as the attachment’s title, and sometimes as the two run together — then take every run of a digit followed by digits, commas or a decimal point, and compare the two collections with collections.Counter. Miss either of the two bare forms and the Q3 totals come out eight tokens apart, which reads as a difference and is not one. Overlap: difflib.SequenceMatcher(None, a, b, autojunk=False).find_longest_match. The reconciliations are arithmetic on figures the reports themselves print. Every count names the file it came from.

What is carried forward

Almost nothing, which was the first surprise. Split each Carolinas report at its blank lines, collapse the runs of whitespace, and keep the blocks of twenty words or more — the prose, not the table cells. Fall 2025 has 19 such paragraphs, and 2 of them appear word for word in Spring 2025. Spring 2026 has 22, and 1 appears word for word in Fall 2025. All three are footnotes. Not one paragraph of body text survives either transition unchanged, so a diff of consecutive periods hands you the whole report rather than a short list.

One sentence comes close to lasting the whole run. Every Carolinas instalment says the wholesale forecast holds two discrete projects totalling nearly 900 MW by 2035 — from the first report, filed 15 May 2025, to the last, dated 9 July 2026. Even that one is not word for word: the 2025 reports write two (2) discrete projects and the 2026 reports drop the numeral. Nothing says whether anybody looked.

What moves and says so

Georgia’s quarterly reports state their own deltas, and the deltas hold. Four stock series — total pipeline and large-load pipeline, each at the 2028/2029 winter and in the mid-2030s — over three transitions is twelve checks: does this quarter’s level equal last quarter’s plus the change it claims? Ten tie exactly. Both misses are 100 MW and both land on the large-load pipeline: 56,800 less a stated 6,000 is 50,800, and Q3 prints 50,900; 24,400 less a stated 3,000 is 21,400, and Q4 prints 21,500. The component flows are looser — Q1’s three sum exactly to its stated net of 7,900 MW, but Q2’s give 4,800 against a stated 4,700, Q3’s −5,959 against −6,000, Q4’s 15,009 against 15,000. That is what rounding produces when flows are quoted to the megawatt and stocks to the hundred; the report’s own summary table is headed approximate MW. From outside it is also indistinguishable from an error, because the unrounded figures are in the trade-secret version.

What moves and does not say so

Every Georgia report measures growth from a fixed point: customer commitments at the 2023 IRP Update filing in October 2023. Q1, Q2 and Q3 give the same figure. Q4 gives a different one.

Twelve megawatts, on a quantity that stopped being able to change in October 2023. Almost certainly rounding, to match the report’s own approximate table. Nothing records that it moved, so anyone holding growth since the 2023 IRP Update as a series now has a step in it that belongs to the report rather than the world.

The Carolinas do the same to a January 2024 filing nobody is amending, and by more than four times as much.

Fifty megawatts, and this one is decidable: the Fall 2025 table breaks that total into 285 + 2,180 + 1,368, which is 3,833. Spring 2025 gives no components, so 3,883 has nothing to check it against. 3,833 carries forward into Spring 2026. Neither report notes that a published figure was replaced.

The denominator moves too. Georgia’s construction bullet sits in the same place in Q3 and in Q4 and answers a different question in each.

6,175 of 7,800 is 79.2 per cent. 7,970 of 11,800 is 67.5 per cent. Read the two as a series and construction coverage looks nearly twelve points worse. It is not; the basis moved from the 2028/2029 winter to the mid-2030s. Q4 does print the near-term pair, 6,200 of 6,900, but in Table 2 rather than in the bullet that states the ratio.

The category moves. In Spring 2025 the Carolinas’ third tier is the 90% Pipeline and the second holds 22 projects that have executed LOIs. By Fall 2025 the same 22 projects at the same 4,669 MW sit in a row called Letter Agreement, and the third tier is the Late-Stage Pipeline. The entry test moved too.

Neither change is announced. Hold 90% Pipeline megawatts as a series and it ends in October 2025 with nothing to tell you.

The authority moves. Georgia’s Q3 and Q4 both open by naming two stipulations. They close naming one each, and not the same one.

Across four quarters that sentence names the 2023 order, then both, then the 2023 order, then only the 2025 one. No quarter says which order the report now answers to.

One of these is simply a slip. Spring 2026 states its headline twice, in the cover letter and in the body.

Table 1 of the same filing gives 4,337 for that cell, and 4,337 is what makes the column add to the 7,946 MW total the table prints; 4,332 gives 7,941. Sealed agreements had been left out in April, so the report was refiled on 9 July 2026. The report itself carries over untouched: the longest run identical in both files is 8,742 characters, and it holds both the 4,332 and the 4,337. What changes is the cover letter, replaced by a shorter one about the missing agreements. The April letter carried the only sentence in the filing that compared Spring 2026 with the report before it — approximately 30% or more than 2,330 MW — and 2,330 appears nowhere in the July version. Neither does Fall 2025 appear in Table 1, whose three columns run from the 2023 filing to Spring 2025 to Spring 2026 and skip the instalment in between. Fall 2025’s own total, 5,610, has no successor anywhere in the operative report.

Being fair about all of this

None of it is improper. A stipulation requiring a report on changes to the pipeline does not require a changelog of the report’s own prose. Definitions get sharper. Rounding a historical figure to match a table labelled approximate is tidying, not revising. Both filers here found their own errors and refiled without being ordered to: Georgia Power wrote that the error had recently come to its attention, and Duke wrote that the agreements were inadvertently omitted.

My guess is that the difficulty falls on whoever holds these quantities as a series. It is a guess: I have not asked anyone who does. I have read the filings, not the desks that use them, and nothing here is evidence about how anybody works. The arithmetic is not a guess. The series is redefined underneath its own history, and nothing in any of these filings records that. The commission asked what had changed in the pipeline and got an accurate answer. Nothing asks the report what it changed about itself.

What to do on Monday

• Divide the flags by the rows. If a change column carries a Y in every cell, the column is broken. 535 ÷ 107 = 5. 530 ÷ 106 = 5. Do that before reading a row.
• Keep one line of restated history per report. Write down every figure it states about a period already closed, with the file it came from. You are not checking the new number; you are checking whether the old one still says what it said.
• Diff the definitions, not only the numbers. Copy the criteria paragraph out of each instalment into one file and read the versions together. A total whose category was renamed is a different total.
• Write the denominator beside the ratio. Georgia’s construction bullet is correct in both quarters; the pair misleads only because the denominator lives in the prose.

The files

• Georgia PSC, Quarterly Large Load Economic Development Reports, public disclosure versions: Q1 2025, Q2 2025, Q3 2025, Q4 2025, Q3 and Q4 corrected, 4 March 2026.
• NCUC E-100 Sub 208A, Semi-Annual Report on Large Load Additions in Advanced Stages of Development: Spring 2025, Fall 2025, Spring 2026, public, Spring 2026 update, public.
• Line numbers refer to the extracted text as saved, not to pages of the PDF. The Carolinas reports are the public versions; the electric service agreements are under seal and were not retrieved. The Spring 2025 and Fall 2025 reports carry a docket date of 9 April 2026 because that is when the Companies re-filed both of them into Sub 208A; the Spring 2025 report was first filed 15 May 2025.