I think claiming '100% positive' without explaining how you detect bitflips is a red flag, because credible evidence looks like ECC error counters and machine check events parsed by mcelog or rasdaemon, reproducible memtest86 failures, or software page checksums that mismatch at crash time.

Ask them to publish raw MCE and ECC dumps with timestamps correlated to crashes, or reproduce the failure with controlled fault injection or persistent checksums, because without that this reads like a hypothesis dressed up as a verdict.