Imagine you're making a smoothie. You need randomness the way you need ice — it doesn't have to be perfectly uniform cubes from a tray; it just has to be cold enough. For years, cryptographers assumed that if your randomness source was good enough to split secrets (secret sharing), it must also be good enough to extract clean, uniform random bits from. This paper proves that assumption wrong. You can make the smoothie with crushed ice, shaved ice, even slightly irregular chunks — and it still works — even though no machine could reshape those chunks into perfect cubes. The committed claim: there exist classes of imperfect (flat, min-entropy ≥ b) randomness sources that support perfectly correct and perfectly private two-out-of-two secret sharing of b-bit secrets, yet from which no deterministic function can extract even a single nearly-uniform bit. This directly answers a question posed by Bosley and Dodis at TCC 2007, and it answers it in the negative. The construction is startlingly compact. For every b ≥ 2, the paper builds flat sources on {0,1}^{b+2} with min-entropy at least b, constructs a sharing scheme producing two (b+1)-bit shares, and proves that every deterministic map from {0,1}^{b+2} to {0,1} lands at statistical distance at least 1/6 from uniform on some source in the class. The proof technique is a parity argument — short, combinatorial, and self-contained. No computational assumptions, no heavy machinery. Where does this sit on the ladder? The prior art is Aggarwal, Chung, Obremski, and Ribeiro (TCC 2022), who resolved the question positively but only for schemes with additional properties — leakage resilience or non-malleability. With those guardrails, extractability was necessary. Without them, it turns out, it isn't. The Bosley-Dodis result (TCC 2007) established the analogous positive result for encryption of long messages. So the landscape is now cleanly partitioned: encryption of long messages requires extractable randomness; vanilla secret sharing does not. The architecture is pure information-theoretic combinatorics. No algorithms run, no computational hardness assumed. The method leans on the structure of flat distributions and the pigeonhole-style constraints that parity arguments provide. This is a foundations paper, not a systems paper — it reshapes what we know is theoretically possible, not what we can build tomorrow. Integrity is strong in the way that mathematical proofs are strong. The result is a constructive existence proof: here is the source class, here is the scheme, here is the extraction lower bound. There's no simulation to overfit, no benchmark to cherry-pick. The validation is the proof itself, and the proof is short enough that a reviewer can check it line by line. The extension results — two-out-of-n and r-out-of-r thresholds with non-extraction bounds of 1/(4n−2) and 1/(2(2^r−1)) respectively — further stress-test the construction. The obvious successor experiment the author did not run: extending the separation to general access structures beyond threshold schemes, or tightening the 1/6 statistical distance bound. The 1/6 is clean but possibly not tight — an optimized construction might push it closer to 1/2. My read: this is a (c) situation — the author is likely saving the general access structure extension for a follow-up, since the threshold generalizations already appear here and the general case is a natural next paper.
