{"reviews":{"source":{"verdict":"pass","summary":"All 23 supplied claim IDs are supported within their declared scope. Live primary passages, public saved-source revision, dates, versions, complete prose, question answers, equations, code comments, calculations, rendered diagrams and editable XML were reviewed. No material unsupported generalizations, omitted claims or private material were found. This is automated source-support review, not independent factual verification; the supplied execution evidence covers synthetic arithmetic only.","issues":[],"claims":[{"id":"C1","verdict":"supported","reason":"Equation 1 and section 2 explicitly support weighted expert outputs and skipping zero-gated computation. [Shazeer et al.](https://arxiv.org/pdf/1701.06538)"},{"id":"C2","verdict":"supported","reason":"The abstract, architecture table and section 2.1 support two-of-eight routing, selected-logit softmax and rounded 47B/13B counts. [Mixtral](https://arxiv.org/html/2401.04088v1)"},{"id":"C3","verdict":"supported","reason":"Equations 3–5 and the adjacent gradient discussion support Gaussian noise, learned softplus scaling and selected-gate derivatives. [Shazeer et al.](https://arxiv.org/pdf/1701.06538)"},{"id":"C4","verdict":"supported","reason":"Equations 1–2 retain full-softmax probabilities for selected experts; singleton renormalization would mathematically give one. [Switch](https://www.jmlr.org/papers/volume23/21-0998/21-0998.pdf)"},{"id":"C5","verdict":"supported","reason":"Section 4 describes self-reinforcing imbalance and separate importance/load objectives; the notebook appropriately avoids inevitability. [Shazeer et al.](https://arxiv.org/pdf/1701.06538)"},{"id":"C6","verdict":"supported","reason":"Equations 4–6 define the stated top-1 loss, hard fractions and differentiable mean probabilities. [Switch](https://www.jmlr.org/papers/volume23/21-0998/21-0998.pdf)"},{"id":"C7","verdict":"supported","reason":"Equation 3 and its following paragraph support capacity scaling and residual bypass for overflow. [Switch](https://www.jmlr.org/papers/volume23/21-0998/21-0998.pdf)"},{"id":"C8","verdict":"supported","reason":"The abstract and computation discussion support dropless block-sparse processing of variable expert workloads, without the broader guarantees explicitly excluded here. [MegaBlocks](https://arxiv.org/pdf/2211.15841)"},{"id":"C9","verdict":"supported","reason":"Section 3.2 and equations 1–2 establish fixed expert budgets and variable token participation. [Expert Choice](https://arxiv.org/pdf/2202.09368)"},{"id":"C10","verdict":"supported","reason":"Sections 3.1–3.2 support intermediate-width segmentation and always-selected shared experts; redundancy reduction is appropriately framed as the authors’ motivation. [DeepSeekMoE](https://arxiv.org/html/2401.06066v1)"},{"id":"C11","verdict":"supported","reason":"Equations 12–20 and section 4.2 support sigmoid affinities, bias-adjusted selection, unbiased normalized gates, sequence balancing and the stated expert/node configuration. [DeepSeek-V3](https://arxiv.org/html/2412.19437v2)"},{"id":"C12","verdict":"supported","reason":"Sections 3.2.1 and 3.4.2 describe dispatch/combine phases and direct point-to-point implementation. [DeepSeek-V3](https://arxiv.org/html/2412.19437v2)"},{"id":"C14","verdict":"supported","reason":"Recalculation confirms both parameter totals, the 27.16% fraction and the 5,637,144,576-parameter sensitivity. Accounting exclusions and runtime limitations are explicit."},{"id":"C15","verdict":"supported","reason":"Selected logits yield weights 0.75/0.25 and output [1,0]. The attached routing render and XML agree on selected branches, inactive experts and weighted-sum dependencies."},{"id":"C16","verdict":"supported","reason":"Code inspection supports stable normalization, synthetic seeded noise, deterministic tie-breaking and separate top-1/top-2 calculations. No model-runtime validation is implied."},{"id":"C17","verdict":"supported","reason":"The stated fractions and means give 0.009344. A sufficiently small positive replacement for the zero probability preserves the strict inequality."},{"id":"C18","verdict":"supported","reason":"Enumerating routes confirms counts [3,2,2,1], rejection of (2,0) at capacity two, and four unused slots at capacity three."},{"id":"C19","verdict":"supported","reason":"Column-wise affinity rankings give participation [2,1,1] with two slots and [0,1,1] with one slot."},{"id":"C20","verdict":"supported","reason":"The binomial counts are 28 and 1,820; four half-width experts equal two coarse-width experts under the explicitly stipulated model."},{"id":"C21","verdict":"supported","reason":"Adjusted scores select experts 2 and 1; original affinities normalize to 0.4615/0.5385. The stated overload sign update also checks."},{"id":"C22","verdict":"supported","reason":"Placement enumeration confirms three remote pairs, rank workloads five/three and 49,152 payload bytes. Rendered arrows and XML correctly show local processing and remote dispatch/return."},{"id":"C23","verdict":"supported","reason":"Dividing stipulated 2nP FLOPs by Pb bytes yields 2n/b. The prose retains the weight-only assumptions and avoids universal latency conclusions."},{"id":"C24","verdict":"supported","reason":"The public-original tool returned the specified discovery revision. Primary dates and versions agree with the bibliography; disclosure separates research reports, illustrative examples and performance verification."}],"package_sha256":"3b490485f6984924acdd37525afb75b15c7184752f143e09921c2681becf3014","reviewed_at":"2026-11-16T18:40:03.808551+01:00"},"visual":{"verdict":"pass","summary":"Reviewed all supplied desktop, 390px and 320px mobile screenshots, scroll-panel positions, question states, diagram sources and ten print pages in the attached manifest. No material visual or visible semantic issues found. Styling follows the approved references; diagrams, arrows and text alternatives agree with the supplied content. Code is syntax-coloured, question rows have rounded borders and badges, and numbered references retain full URLs. Mobile overflow stays within panels, with remaining content demonstrated in scrolled views. Print content remains readable and complete. Recorded checks report successful keyboard navigation, copying, contents links and downloads. This is a visual review, not factual certification or an independent rerun of browser checks.","issues":[],"claims":[],"package_sha256":"3b490485f6984924acdd37525afb75b15c7184752f143e09921c2681becf3014","reviewed_at":"2026-11-16T18:40:03.808610+01:00"}},"checks":{"package_sha256":"3b490485f6984924acdd37525afb75b15c7184752f143e09921c2681becf3014","passed":true,"widths":[1440,390,320],"code_checked":true,"code":[{"language":"python","syntax":"passed","sha256":"c6a6791111563b33e2d7380e67b0322a8c83fe18fb0d2a0d42ce2412f98c8461"}],"execution":[{"file":"lab.py","sha256":"22a5cebcca256fc25621b1ee21d1f98c0c5a80f0626720fdd5599d76b3a3b250","returncode":0,"output":"Synthetic top-2 counts: [3, 2, 2, 1]\nSelection rates: [0.75, 0.5, 0.5, 0.25]\nAssignment shares: [0.375, 0.25, 0.25, 0.125]\nCapacity-two admission: [2, 2, 2, 1] [(2, 0)]\nIllustrative noisy counts: [3, 2, 2, 1]\nSeparate top-1 counterexample: [0.9, 0.1] [0.45899999999999996, 0.541] 0.009344\nExpert-choice participation: [0, 1, 1]\nUnbiased weights after biased selection: [(2, 0.46153846153846156), (1, 0.5384615384615385)]\nRemote assignments and payload bytes: [(0, 3), (2, 0), (3, 1)] 49152\nSynthetic stored / selected-path parameters: 46433042432 12610174976\nSelected-path fraction: 0.2715776161871641\n","environment":"CPython 3.12 isolated container; no network, no host mounts; 128 MiB; 1 CPU"}],"drawio":["routing.drawio","parallel.drawio"],"print":"rendered"},"published_at":"2026-11-16T17:40:16.135Z"}