# Review: cross-reused replication vs. new laboratories in a lab-or-method artifact model **Bottom line:** The claim is correct as an existence claim, with caveats. "Same leading-order gain" holds for the ε-exponent and for the leading term of log-evidence, but the exact leading coefficient of the Bayes factor differs by a factor of 2. The underlying mathematics is a direct application of standard results. ## 1. Precise restatement with hidden assumptions **Setup.** - There is a bipartite graph G with laboratories L on one side and methods M on the other. Edges are experiments, and every observed experiment succeeded. - H predicts success with probability 1. - Under the rival R: - each vertex independently carries a *persistent* artifact with the *same* probability ε; - an experiment succeeds iff its laboratory or its method carries an artifact. - Evidence is measured by the Bayes factor BF = P(all succeed | H) / P(all succeed | R), in the limit ε → 0. - "Leading-order gain" means the change in the exponent of ε. Equivalently, it is the change in the leading term of log BF. **Claim.** There exist a graph G and an edge e such that: - e joins an *existing* laboratory to an *existing* method; - adding e raises the exponent by 1, exactly as adding an isolated new-lab/new-method edge does; - there is also a way of adding arbitrarily many new laboratories (all attached to one existing method) that leaves the exponent unchanged. The claim is existential. It does not say that cross-reuse always helps, or that new labs never help. ## 2. Checking the argument and re-deriving it **Likelihood under R.** P(all succeed | R) is the probability that the artifact-bearing set S is a vertex cover: P = Σ over covers S of ε^|S| (1−ε)^(n−|S|) = c·ε^τ + O(ε^(τ+1)). Here τ is the minimum cover size and c is the number of minimum covers. Since P(data | H) = 1, we get BF = c⁻¹ ε^(−τ)(1 + O(ε)), and so log BF = τ log(1/ε) − log c + O(ε). By König's theorem, τ equals the maximum matching size ν in a bipartite graph. All of this is correct. **Original graph.** The edges are A–x, A–y, B–z, C–z. - The matching {A–x, B–z} gives ν = 2. - A cover of size 2 must cover {Ax, Ay}, which needs A or both of {x, y}. It must also cover {Bz, Cz}, which needs z or both of {B, C}. - The only cover of size 2 is therefore {A, z}. So τ = 2 and c = 1. **After adding B–x.** B and x both already exist, so this edge introduces neither a new laboratory nor a new method. - The matching {A–y, B–x, C–z} has size 3, so τ = 3. - Enumerating the minimum covers gives {A, B, z}, {A, x, z}, {A, B, C}, and {x, y, z}. So c = 4. **Isolated new edge D–w instead.** - τ = 3. - The minimum covers are {A, z, D} and {A, z, w}, so c = 2. **Comparison.** - The cross-reused edge multiplies BF by about 1/(4ε). - The isolated edge multiplies BF by about 1/(2ε). - The exponent gain is identical (+1), and the leading log-evidence gain log(1/ε) is identical. - The constants differ: the isolated edge gives about twice the Bayes-factor gain. So "same leading-order gain" holds in exponent or log terms, but not for the exact leading coefficient. **New labs using z.** Add labs D₁, …, D_k, each using method z. - The star at z still has cover {z}, so τ stays 2 and c stays 1. - The exact factor is P(z or all leaves have artifacts) = ε + (1−ε)ε^(k+2). This changes only at higher order. - So there is no leading-order gain for any k. This is correct. ## 3. Attempts to break it **Where the result is fragile.** - **Dependence on configuration.** One new laboratory using *x* (D–x) would raise ν to 3 via A–y, D–x, B–z. "New labs give no gain" therefore depends on the specific attachment to z. The existential "can" covers this. - **Dependence on the ε → 0 limit.** For moderate ε, the higher-order terms and the constants c matter, and rankings can shift. For example, the factor of 4 versus 2 matters when ε is not very small. - **Dependence on equal artifact rates.** If labs and methods had different rates ε_L and ε_M, minimum covers would be replaced by minimum-*weight* covers. The "matching size" characterization would then change, because König's theorem is about cardinality. **What holds in general.** - A single added edge changes ν by 0 or 1. - An isolated new edge always adds exactly 1. **Counterexamples.** I found none to the existential claim as stated. ## 4. Novelty The mathematics is standard: - König's theorem; - the small-p expansion of the probability that a random vertex set is a cover, which is the vertex-cover reliability polynomial. The philosophical framing echoes the variety-of-evidence and robustness literatures. The specific "augmenting-structure" reading of evidential novelty may be a new framing, but the claim itself is an illustrative instance of known combinatorics. ## 5. Fairness The claim is mostly faithful to the text. However: - it leaves implicit the ε → 0 limit, the equal artifact rates, and that H predicts certain success; - "same leading-order evidential gain" is accurate for the exponent and log-evidence, but would mislead a reader who takes it to mean the leading asymptotic term of the Bayes factor, which differs by a factor of 2. {"verdict": "correct with caveats", "novelty": "special case or variant of known results", "fair_summary": "partly", "confidence": 4, "one_line": "The example checks out (τ goes 2→3 via the cross-reused edge B–x, and stays 2 under any number of new labs on z), but 'same leading-order gain' holds only for the ε-exponent/log-evidence (constants c=4 vs 2), and it relies on ε→0 and equal artifact rates."}