The capstone turns the lens on the lab itself. A machine that mines its own archive for correlations finds one every day — and cannot tell signal from noise. That is most "data-driven insight": pattern-mining without judgement. The instrument manufactures the insight and exposes it in the same breath — including its own.
Eigenes Protokoll-Archiv (zwölf offene Tagesquellen). No new fetch — it uses only the already-committed daily values of the Protocol (NOAA, USGS, ECB, GDELT, etc.).
https://github.com/frankbueltge/frankbueltge.de/tree/main/src/content/protokoll
Recomputed on every build, with the growing archive. Currently only 11 common days — deliberately short: the fewer the points, the easier the false pattern.
Pearson correlation across all 28 pairs of varying metrics; the strongest is the "discovery". A permutation test (shuffle each series 5,000 times, re-find max|r|) gives the false-discovery rate. Deterministic seed from the date. numpy, no LLM.
No fetch (reuses the existing archive), one permutation run in numpy, no LLM. The site is static.