I recorded raw bytes from 58 hardware entropy sources on a Mac Mini M4 (clock jitter, DRAM timing, PLL oscillators, interrupt scheduling, a USB QRNG, etc.) using an open-source framework I built called OpenEntropy. Then I embedded the byte windows with text-embedding-3-large and compared them.
19 sources from 10 different hardware categories landed in the same cluster. After removing the dominant shared factor, they stayed clustered. Under stress tests, sources like dram_row_buffer and page_fault_timing moved together while the USB quantum RNG barely moved at all.
I haven't tested whether SHA-256/ChaCha20 conditioning removes the correlations (it probably does), and this is one machine. But per-stream tests like NIST SP 800-22 wouldn't catch this since they don't look across sources.