WEBVTT

1
00:00:00.000 --> 00:00:04.500
Case 2 · Real data: the same fact counted more than once. Foreign direct investment statistics, SDMX cube, accessed over MCP.

2
00:00:04.500 --> 00:00:09.000
Question: what was the notified US foreign direct investment in Korea in 2020?

3
00:00:09.000 --> 00:00:27.000
Summed without the standard, 35 rows are added: duplicate reports across source files, detail and total rows counted together, and rows with status Z or M. 21,208,000 thousand USD, with no error.

4
00:00:27.500 --> 00:00:44.000
With the manifest's conditions — deduplicate source files, industry = ALL, investment type = ALL, status A or C — 2 rows remain: 5,302,000 thousand USD.

5
00:00:44.500 --> 00:00:56.000
The unconstrained sum is 4.0 times the correct value. The needed conditions are not in the data; they are recorded in the manifest. The "without" query is a representative query, not one written by an LLM.

6
00:00:56.500 --> 00:01:04.000
A wrong total raises no error. The conditions that separate duplicates, totals and observation status must be recorded with the data in machine-readable form.