Imagine you're hosting a dinner party where four friends each need to receive a private envelope, but the only way to pass information is by placing cards on the table edges between pairs of guests. Each guest can see only the cards touching their seat. Your challenge: how many extra "dummy" cards — pure randomness — do you need so that every guest learns only their own secret and nothing about anyone else's? That's the core mechanism here. More dummy cards means more privacy, but randomness is expensive. This paper finds the exact minimum. The committed claim: for four binary secrets on a complete graph, the minimum number of random states required is exactly 1, 2, 3, or 4, depending on the correlation structure of the permitted secret combinations. The paper fully classifies which correlation structures land in which bucket. This extends the complete characterization by Anilkumar et al. (2024) from three secrets to four — a jump that is combinatorially much harder, since the number of possible correlation structures explodes. The architecture is pure combinatorial information theory. No algorithms run on hardware, no neural nets, no optimization solvers in the traditional sense. The core tools are Shannon entropy inequalities, linear programming over entropy cones, and exhaustive computational enumeration of all possible access structures on four-party graphs. The code on GitHub automates the LP-based lower bound proofs and the exhaustive case analysis — this is a math paper with computational verification, not a systems paper. Integrity is strong for this type of work. The results are mathematical theorems with proofs, not empirical benchmarks. The computational enumeration covers all cases and the code is released. The proofs in the appendix are self-contained. The main vulnerability isn't cherry-picking — it's whether the proof techniques generalize beyond four parties. The authors are honest: they fully solve 3 and 4 parties but acknowledge that the general case remains open. The ladder here is Anilkumar et al. (2024), which solved the three-party case completely. This paper extends that baseline to four parties and also provides partial results for arbitrary numbers of secrets on general graphs, specifically characterizing when a single random bit suffices on any graph. The jump from 3 to 4 is genuine progress — the combinatorial complexity increases substantially and the proof techniques had to be extended. The natural milestone is five parties on a complete graph, then general n. The number of possible secret correlation structures grows super-exponentially with n, so brute-force enumeration will not scale. New structural theorems — not just more compute — will be needed. The gap between 4 and general n is where the real theoretical difficulty lives. The obvious experiment not run is the five-party case. The honest read: the computational enumeration already required significant effort at four parties. Five parties likely pushes the LP-based approach past feasible enumeration without new structural insights to prune the search space. This isn't compute they're saving for a sequel — it's a genuinely harder mathematical problem that requires new ideas.
