We prove that Max-3-Cut is NP-hard to approximate within a factor of $27/29+\varepsilon \approx 0.931$. This improves the factor $16/17 \approx 0.941$ of Austrin, O'Donnell, Tan and Wright, the best NP-hardness result for Max-3-Cut since 2012. (Their result, unlike ours, also holds for 3-colourable graphs.) We also improve the NP-hardness of Max-$k$-Cut for every $k$ from 4 to 20; for instance, we obtain $45/47 \approx 0.957$ for Max-4-Cut and $33/34 \approx 0.971$ for Max-5-Cut, where the previous factors were 0.987 and 120/121. Our reductions are gadget reductions from Håstad's linear equations over $\mathbb{Z}_3$, with gadgets that have at most six auxiliary vertices. The main new idea is a negation trick: since the negation of an assignment is again an assignment, equations with right-hand sides $t$ and $-t$ can be handled together, and this makes the resulting predicates symmetric under all permutations of the colours. We complement it with a general soundness criterion for such reductions: in effect, the best strategy of the adversary is to use a single colour pattern everywhere. This criterion is the strongest analysis that uses Håstad's theorem as a black box, and, using LP duality, we show that under it no gadget reduction from these equations can beat $27/29$ for Max-3-Cut. Finally, we observe that a classical reduction from Max-3-Cut to Max-$k$-Cut, applied more carefully, already gives $82/85 \approx 0.965$ for Max-4-Cut on 4-colourable graphs, improving the previously stated $85/86$.
Numbers computed from trained models, such as an example's difficulty or a model's benchmark lead, change with seed and prompt. We ask when such measurements can be corrected or compared. Noise shrinks correlations, and the textbook correction divides by the square root of two estimated reliabilities. On two CIFAR-100 difficulty scores, it returned correlations above one in 70 of 100 classes. We prove that its mean squared error is infinite even when the correlation and the reliabilities use separate seeds. Our rule corrects only when both reliabilities pass a significance test, then divides by their upper confidence bounds. Its output is always bounded, and its coverage guarantees assume Gaussian noise. It helps once reliability is well estimated, more than halving the error with 500 simulated units. Prompt settings can reverse rankings. On ARC, Qwen2-1.5B trails Llama-3.2-3B by 3.8 points with no examples in the prompt and leads by 16.6 with two. Across four benchmarks, averaging the two settings names one winner for 15 of 16 such reversals. We compare models within each setting, controlling family-wise error asymptotically in the number of items for a fixed prompt grid. Aligning embedding coordinates across seeds raises held-out per-dimension reliability from 0.003 to 0.813.