## 1. Precise restatement and assumptions The claim is correct **as a statement about asymptotic evidence exponents**, not necessarily about equal leading coefficients. Let \(G=(L\sqcup M,E)\) be a finite bipartite graph. Assume: - Each vertex represents a laboratory or method. - Under the artifact rival \(R\), every vertex independently carries a persistent artifact with the same known probability \(0<\varepsilon<1\). - An experiment succeeds if and only if at least one endpoint carries an artifact. There is no additional experimental noise. - Under \(H\), every experiment succeeds with probability one. - The evidence \(D_G\) consists solely of observing success on every edge. The experimental design is fixed, or its selection contributes no additional likelihood ratio. - The asymptotic limit is \(\varepsilon\to0\), ordinarily with the graph fixed. - Repeating an already-observed laboratory–method pairing supplies no further information in this deterministic persistent-artifact model. Define \[ \operatorname{BF}(G)=\frac{P(D_G\mid H)}{P(D_G\mid R)} =\frac1{P(D_G\mid R)} \] and its evidential exponent \[ k(G)=\lim_{\varepsilon\to0} \frac{\log\operatorname{BF}(G)}{\log(1/\varepsilon)}. \] The precise claim is existential: 1. There is a graph \(G_0\) and an additional edge between existing vertices such that \(k\) increases by one. 2. Adding an isolated edge with two new endpoints also increases \(k\) by one. 3. For every number \(n\) of new laboratories, there is an extension of this same \(G_0\), using those laboratories only with method \(z\), that does not increase \(k\). This does **not** assert that every cross-reused experiment increases evidence, that new laboratories never help, or that the two beneficial extensions have asymptotically identical Bayes factors. ## 2. Checking the argument and independently deriving the result ### The general likelihood calculation Let \(S\) be the set of artifact-bearing vertices, and let \(N\) be the total number of vertices. All experiments succeed exactly when \(S\) is a vertex cover. Consequently, \[ P(D_G\mid R) =\sum_{\substack{S\subseteq V(G)\\S\text{ is a vertex cover}}} \varepsilon^{|S|}(1-\varepsilon)^{N-|S|}. \] If the minimum cover size is \(\tau\), and there are \(c\) minimum covers, their contribution is \[ c\varepsilon^\tau(1-\varepsilon)^{N-\tau} =c\varepsilon^\tau+O(\varepsilon^{\tau+1}). \] Every other cover contributes only at order \(\varepsilon^{\tau+1}\) or smaller. Thus \[ P(D_G\mid R)=c\varepsilon^\tau+O(\varepsilon^{\tau+1}), \qquad \operatorname{BF}(G)=\frac1c\varepsilon^{-\tau}(1+O(\varepsilon)). \] That step is correct. For bipartite graphs, **Kőnig’s theorem** gives \[ \tau(G)=\nu(G), \] where \(\nu\) is the maximum matching size. Hence \(k(G)=\nu(G)\). ### Checking the particular graphs The original edges are \[ A\!-\!x,\quad A\!-\!y,\quad B\!-\!z,\quad C\!-\!z. \] - The matching \(\{A\!-\!x,B\!-\!z\}\) has size two. - The cover \(\{A,z\}\) has size two, ruling out a larger matching. Therefore \(\tau=\nu=2\). After adding \(B\!-\!x\), the three edges \[ A\!-\!y,\quad B\!-\!x,\quad C\!-\!z \] form a matching. There are only three vertices on either side, so the maximum matching size is exactly three. An isolated new edge also raises the matching size by exactly one: matching sizes add across disconnected components. Finally, adding any number of new laboratories connected only to \(z\) leaves \(\{A,z\}\) as a cover. The original size-two matching remains, so the maximum matching size stays exactly two. These arguments completely establish the claimed exponent changes. ### Independent probability calculation We can verify the construction directly, without relying on vertex-cover or matching theory. A two-leaf star succeeds when its central vertex is artifact-bearing, or when it is clean and both leaves are artifact-bearing. Its success probability is \[ a=\varepsilon+(1-\varepsilon)\varepsilon^2. \] The two initial stars involve disjoint vertices, so \[ P_0=a^2. \] After adding \(B\!-\!x\), subtract those initial successes for which both \(B\) and \(x\) are clean. In that case, the old edges force both \(A\) and \(z\) to carry artifacts; \(C\) and \(y\) are unrestricted. Thus \[ \begin{aligned} P_{\times} &=a^2-\varepsilon^2(1-\varepsilon)^2\\ &=4\varepsilon^3-2\varepsilon^4-2\varepsilon^5+\varepsilon^6. \end{aligned} \] Therefore \(P_{\times}\sim4\varepsilon^3\), confirming exponent three. An isolated new experiment succeeds with probability \(2\varepsilon-\varepsilon^2\), giving \[ P_{\mathrm{isolated}}=P_0(2\varepsilon-\varepsilon^2) \sim2\varepsilon^3. \] For \(n\) new laboratories using only \(z\), \[ P_n= \bigl[\varepsilon+(1-\varepsilon)\varepsilon^2\bigr] \bigl[\varepsilon+(1-\varepsilon)\varepsilon^{n+2}\bigr] \sim\varepsilon^2. \] This independently verifies every exponent claimed in the text. ## 3. Attempts to break the claim ### A. “Same leading-order gain” must not mean identical leading terms The Bayes-factor gains relative to the initial graph satisfy \[ \frac{\operatorname{BF}(G_\times)}{\operatorname{BF}(G_0)} =\frac{P_0}{P_\times} \sim\frac1{4\varepsilon}, \] whereas \[ \frac{\operatorname{BF}(G_{\mathrm{isolated}})} {\operatorname{BF}(G_0)} =\frac1{2\varepsilon-\varepsilon^2} \sim\frac1{2\varepsilon}. \] They have the **same exponent**, but different leading coefficients. The cross-reused experiment supplies asymptotically half the multiplicative Bayes-factor gain of the isolated experiment. Equivalently, their log-evidence gains are \[ \log(1/\varepsilon)-\log4+O(\varepsilon) \quad\text{and}\quad \log(1/\varepsilon)-\log2+O(\varepsilon). \] Their leading divergent terms agree. ### B. Additional laboratories do provide some finite-\(\varepsilon\) evidence For \(n\ge1\), \[ \frac{\operatorname{BF}(G_n)}{\operatorname{BF}(G_0)} = \frac{1+(1-\varepsilon)\varepsilon} {1+(1-\varepsilon)\varepsilon^{n+1}} >1. \] Thus “no gain” would be false literally. What is true is **no exponent gain**. Indeed, \[ 1\le \frac{\operatorname{BF}(G_n)}{\operatorname{BF}(G_0)} \le1+(1-\varepsilon)\varepsilon, \] uniformly in \(n\). Even arbitrarily many such laboratories produce a gain tending to one as \(\varepsilon\to0\). ### C. The quantifiers and attachment pattern matter Neither reuse nor laboratory novelty alone determines the result. - **Reuse need not help:** start with \(A\!-\!x,B\!-\!y\), and add \(A\!-\!y\). All endpoints are reused, but the maximum matching remains size two. - **A new laboratory can help without a new method:** start with \(A\!-\!x,A\!-\!y\), and add \(B\!-\!x\). The maximum matching rises from one to two. The supporting construction establishes an existence claim, not a universal rule against laboratory novelty. ### D. The artifact assumptions are essential The result depends on persistent endpoint-level artifacts, independent equal-probability artifact assignments, deterministic success conditional on those artifacts, and the specified small-\(\varepsilon\) limit. For example, if vertex artifact probabilities scale as different powers \(\varepsilon^{w_v}\), the exponent becomes a **minimum-weight vertex-cover cost**, not the ordinary maximum matching size. Additional noise or correlated artifacts can change the result further. There is no unproved mathematical step within the stated model. ## 4. Novelty The mathematical mechanism is a straightforward application of established results: - **Kőnig’s theorem:** minimum vertex-cover size equals maximum matching size in bipartite graphs. - **Berge’s augmenting-path theorem:** a matching is maximum exactly when no augmenting path exists. - **Rare-event analysis of monotone Boolean events:** the lowest-order probability is determined by the smallest sets of rare causes sufficient to produce the event. There is also an exact connection to the **independence polynomial**. The clean vertices must form an independent set, so \[ P(D_G\mid R) =\varepsilon^N I_G\!\left(\frac{1-\varepsilon}{\varepsilon}\right), \] where \(I_G(t)\) counts independent sets by size. Its highest-degree terms reproduce the minimum-cover exponent and coefficient. Thus this is **not a new graph-theoretic or asymptotic-probability theorem**. The interpretation as an example about replication and evidential diversity may be useful or original in presentation; the supplied material does not establish priority for that application. ## 5. Fairness of the one-sentence claim The sentence fairly captures the construction **if “leading-order evidential gain” means an increase in the evidential exponent**, as the supporting text intends. It is less precise than it should be because: - “Same leading-order” can suggest equal leading asymptotic terms, which the calculations disprove. - “No such gain” must mean no exponent increase, not no additional evidence at all. A more exact formulation would be: > In this independent persistent-artifact model, an experiment using an existing laboratory and method can increase the small-artifact-probability Bayes-factor exponent by the same amount as an experiment introducing both, while arbitrarily many new laboratories using one shared method need not increase that exponent. {"verdict":"correct with caveats","novelty":"special case or variant of known results","fair_summary":"partly","confidence":5,"one_line":"The construction rigorously establishes equal increases in evidential exponent, not equal leading Bayes-factor coefficients, using classical vertex-cover and matching results."}