One row reclassified, two evidence packs pinned, sixty articles hand-coded.
A primary-source audit pass caught one misclassified data-center line, pinned the underlying citations for every row, and built the calibration set that the next release grades a real scorer against.
The founding paper logged twelve U.S. AI data-center cancellations, delays, or scale-backs. One of those rows — Meta, 14 May 2025, 0.30 gigawatts — was flagged for re-reading. The original logging treated Meta's Q1 2025 earnings remarks as a U.S. data-center scale-back. A careful re-read of CFO Susan Li's call commentary reclassifies the line. What Meta described was not a retreat from data centers but a capex reweighting into AI within an existing budget envelope. Same dollars, different bucket. The row stays in the ledger with the reclassification note; a future release will either drop it or split the scale-back tag into net retreat versus category reshuffle.
Two evidence packs were pinned alongside the reclassification. The first attempts a primary-source pin on all twelve logged U.S. data-center events; the second does the same for the fifteen Chinese capex rows. The Alibaba FY25 Q3 call, the Financial Times ByteDance dispatch, the Ministry of Science and Technology East-Data-West-Compute Phase II release, and the provincial 14th Five-Year-Plan supplements are now traceable from the page back to a primary citation in a single hop.
The biggest piece of housekeeping is the calibration set itself: sixty hand-coded articles — thirty U.S., thirty Chinese — spanning six blocks. On the U.S. side: peak bull (2023-Q1 through 2024-Q2), normalization (2024-Q3 through 2025-Q2), and anxiety (2025-Q3 through 2026-Q2). On the Chinese side: slow-build (2023 through 2024), confidence (2025), and ascendant (2026). Each article carries a hand-coded sentiment score on a −1 to +1 scale with a one-sentence justification. This is the set that the lexicon scorer is graded against and that the LLM scorer will be graded against.
Technical detail
- Meta 14 May 2025 line: row tagged in the page caption. Disposition pending in a future release — either drop the row or split the scale-back status into net retreat versus category reshuffle.
- Data-center evidence pack: primary-source pin attempted across all twelve logged events.
- Chinese capex evidence pack: Alibaba FY25 Q3 earnings transcript, Financial Times ByteDance dispatch URL, Ministry of Science and Technology East-Data-West-Compute Phase II Xinhua release, and provincial 14th Five-Year-Plan supplement PDFs.
- Calibration set: 60 articles, six blocks (three U.S., three Chinese), hand-coded on a −1 to +1 scale. Feeds the next-release reliability diagram.
What this opens. The reclassification pattern matters more than the single row. The original logging assumed any “capex scale-back” phrase was a data-center retreat; the Meta case shows that hyperscalers also reshuffle within an envelope. The status tagging needs to distinguish those cleanly, or the ledger will read more bearish than reality. The calibration set is the asset that turns the next release from a narrative reading into a scored model.