Page and pipeline, reconciled.
An audit found that some corrections we had announced were missing from the underlying analysis code. This update closes the gap and restores the pipeline as the single source of truth.
Three weeks after the source-pinning revision, we audited the public page against the underlying analysis pipeline and found a quiet drift: some of the corrections we had announced were only applied to the page's display table, not to the code file that everything downstream reads from. Three event rows still showed stale citations. One inductive event we had said was removed was still in the corpus. This update closes the gap.
On the page itself: three source URLs in the historical-events table were updated to match what we had already claimed — a South China Morning Post report from June 2024, The Information's March 2025 dispatch on travel curbs at DeepSeek, and the Bloomberg dispatch from May 2026 that anchors the founding press citation for the index. In the underlying code: an inductively projected pre-travel-review event at Tencent — previously flagged as over-extrapolated and supposedly removed — was in fact still in the events file. It is now actually removed. The other three URLs, the audit confirmed, were already correctly pinned in the code; the page was the one that had drifted.
The disposition matters more than the individual edits. Before this fix, the analysis pipeline could be re-run at any time and would silently revert page-only corrections. Page and code now agree, and a fresh run reproduces the page's claims rather than overwriting them. That's the precondition for an automated research loop to be trustworthy at all.
Technical detail
- Page corrections: three source-URL rows in Table 1 updated to South China Morning Post (Zhuang & Mai, 2024-06-06), The Information (2025-03-14), and Bloomberg (2026-05-26) respectively.
- Code correction: the inductively projected Tencent pre-travel-review event removed from the historical-events file (Tencent was not named in the founding Bloomberg dispatch; the previous revision flagged the extrapolation but the file was stale).
- Verification: re-running the pipeline post-fix produces an analysis manifest identical to the page's claims; no silent reversion.
What this opens. The pattern generalizes. Every paper now ships a run manifest with code, data, and configuration hashes alongside its metric values, so future drift between code and page is detectable rather than silent. That is the substrate the next set of automated checks will be built on.