The core statistical assertion — roughly 2 million retail investors in Trump meme coins collectively lost approximately $4 billion — is substantially supported by multiple independent sources. CryptoRank and Chainalysis, two independent blockchain analytics firms, both report approximately 2 million wallets underwater and combined losses from $TRUMP and $MELANIA of $4.3 billion, figures formally cited in a Senate Banking Committee minority letter dated April 8, 2026 and corroborated by wire and specialist financial reporting . Steve Rattner's own May 1, 2026 X post — the direct source Psaki is paraphrasing — states 'over $4 billion for over 2 million retail investors' , which Psaki rounds to '$4 billion' and '2 million.' The sole identifiable inaccuracy is directional but immaterial: Psaki's '$4 billion' is a rounding-down of the $4.3 billion figure, which technically understates rather than overstates the losses. Per the MOSTLY_TRUE standard, the core assertion is substantially correct but contains an identifiable imprecision — the $4B figure is $300 million short of the sourced $4.3B — that does not reverse the directional meaning; a corrected figure would only strengthen, not undermine, the claim's point. A secondary definitional precision issue exists: the '2 million' figure refers specifically to wallets 'currently underwater,' not all wallets that ever bought the coins, and combines $TRUMP and $MELANIA holders; Psaki's phrasing 'invested in the coins' is consistent with both the combined-token scope and the 'underwater' qualifier. The NYT/Chainalysis single-token analysis ($2B loss, 813,294 wallets for $TRUMP alone) represents a narrower scope and does not contradict the broader combined-token figure; it establishes that the $4B/$2M figure is the more comprehensive and current estimate. No corrections have been issued. MOSTLY_TRUE is the appropriate verdict: the core assertion is substantially correct, the rounding imprecision is minor and unidirectional (understating losses), and the corrected figure would not change a reasonable viewer's interpretation of the overall point (per the MOSTLY_TRUE boundary test).