**Verdict:** The claim is correct. I checked every step independently and confirmed it with an exact calculation for r = 2. Its unstated conditions are that the theory set grows, N = M^r → ∞, and that the prior is uniform. ## 1. Precise restatement Fix r ≥ 2. For every ε > 0 there is a finite theory set Θ with |Θ| = N, and two experiments P and Q on Θ, such that: - **(a) Agreement on small menus.** For every subset A ⊆ Θ with |A| ≤ r, the restricted experiments {P_θ : θ ∈ A} and {Q_θ : θ ∈ A} are Blackwell-equivalent. Equivalently, the two experiments are equally valuable in every decision problem whose prior is supported on at most r theories, with arbitrary actions and losses. - **(b) Divergence on the whole set.** Under the uniform prior on Θ, the optimal probability of identifying the true theory is ≥ 1 − ε under P and ≤ ε under Q. Hidden assumptions: - N must grow (N = M^r with M → ∞); "approaching" refers to this sequence. - The uniform prior is used for the identification probability. - The construction uses a special symmetric "random subset containing the truth" family of experiments. ## 2. Checking the steps and re-deriving **Identification probability.** Write p_k = Pr(K = k). The likelihood of reporting S is P(S | θ) = p_{|S|} / C(N−1, |S|−1) · 1[θ ∈ S]. Under the uniform prior, the posterior given S is uniform on S. The maximal posterior is therefore 1/|S|, and the optimal identification probability is E[1/K]. ✓ **Sufficiency for a menu A with |A| = s.** For θ ∈ A, the likelihood factorizes as 1[θ ∈ S∩A] · h(S), with h(S) not depending on θ. So T = S∩A is sufficient on A. Fix a set t with θ ∈ t ⊆ A and |t| = j. Then P(T = t | θ) = C(N−s, K−j) / C(N−1, K−1) = c · (K−1)(K−2)⋯(K−j+1) · (N−K)(N−K−1)⋯(N−K−s+j+1), where c = (N−s)!/(N−1)!. This is a polynomial in K of degree (j−1) + (s−j) = s−1. The falling-factorial form also gives the correct value 0 at the boundary values K = 1 and K = N. ✓ So if the moments of K match up to degree r−1, then for every s ≤ r the law of T agrees for each θ ∈ A. Each experiment garbles to T, and T can be garbled back via the θ-free kernel S | T. Hence the restricted experiments are Blackwell-equivalent. ✓ **Moment matching.** The weights w_j are the divided-difference weights on r+1 nodes. They satisfy Σ w_j f(x_j) = f[x_0, …, x_r], which is 0 whenever deg f ≤ r−1. - The case f = 1 shows the positive and negative parts have equal mass. - Normalizing each part gives two probability distributions with matching moments through degree r−1. - All nodes M^j lie in {1, …, N}. ✓ **Asymptotics.** Let E = r(r+1)/2. - |w_0| ≈ M^{−E}. - |w_1| = 1 / [(M−1) ∏_{l≥2}(M^l − M)] ≈ M^{−E}, and |w_1|/|w_0| → 1. - For j ≥ 2, |w_j| ≈ M^{−E − j(j−1)/2}, which is smaller by at least a factor of 1/M. - The signs are (−1)^r for w_0 and (−1)^{r−1} for w_1, so they are opposite. Consequences: - One distribution has mass 1 − O(1/M) at K = 1, so E[1/K] → 1. - The other is supported on {M, M^2, …}, so E[1/K] ≤ 1/M → 0. ✓ ## 3. Attempting to break it: explicit check for r = 2 Take nodes 1, M, M², so N = M². The weights are: - w_0 = 1/((M−1)²(M+1)) - w_1 = −1/(M(M−1)²) - w_2 = 1/(M(M−1)²(M+1)) **Mass balance.** w_0 + w_2 = (M+1)/(M(M−1)²(M+1)) = |w_1|. ✓ **First moment.** w_0 + M²·w_2 = 1/(M−1)², and M·|w_1| = 1/(M−1)². ✓ **The two distributions.** - P puts mass M/(M+1) on K = 1 and mass 1/(M+1) on K = M². Its mean is M. - Q is a point mass at K = M. Its mean is also M. For s = 2 menus the law of T is affine in K, so it matches. The s = 1 case is trivial. **Identification probabilities.** E_P[1/K] = M/(M+1) + 1/(M²(M+1)) → 1, while E_Q[1/K] = 1/M → 0. **Edge cases checked:** - Boundary values of K (1 and N) are handled by the falling-factorial form. - Both parts of the weight vector are nonempty. - The case r = 1 is excluded, and would be trivial anyway. I found no counterexample or gap. ## 4. Novelty The phenomenon that agreement on small subproblems does not determine full informativeness has close relatives: - Experiments can agree on all dichotomies without being equivalent overall. - Moment-matching constructions are standard in lower-bound arguments in statistics and property testing. This specific divided-difference construction, framed for theory choice, looks like a neat variant rather than a fundamentally new phenomenon. ## 5. Fairness of the one-sentence claim The summary is fair. "Approaching" signals the sequence N → ∞. The unstated uniform prior and the special experiment family are standard for an existence claim of this kind. {"verdict": "correct", "novelty": "special case or variant of known results", "fair_summary": "yes", "confidence": 4, "one_line": "The moment-matching construction is valid: restriction to any ≤r-theory menu yields identical sufficient-statistic laws (degree ≤ r−1 polynomials in K), while E[1/K] tends to 1 versus 0 as N = M^r grows."}