{"reviews":{"source":{"verdict":"pass","summary":"Reviewed all 20 claims, complete prose, question answers, equations, code/comments, supplied rendered diagram and editable XML against live primary-source passages and the cited ResearchNode public original. All are supported within their declared scope. Dates, versions and saved revision match. No material omissions, unsupported generalizations or private evidence found. The supplied isolated execution supports only the synthetic lab, not model or deployment performance. This is automated source-support review, not independent factual verification.","issues":[],"claims":[{"id":"C1","verdict":"supported","reason":"The explicit allocations total 16 bytes; conversions give 1,120 GB and 1,043.081 GiB. The frozen-base subtotal follows under the stated exclusions."},{"id":"C2","verdict":"supported","reason":"The qualified distinction between intrinsic dimension and matrix rank agrees with the abstract and Section 3 of the [ACL paper](https://aclanthology.org/2021.acl-long.568.pdf)."},{"id":"C3","verdict":"supported","reason":"Shapes, freezing, Gaussian/zero initialization, scaling and merging match Section 4.1 of [LoRA v2](https://arxiv.org/pdf/2106.09685)."},{"id":"C4","verdict":"supported","reason":"All three gradients and the merge identity follow by differentiation and distributivity. Counts and ratio are correct. Rendered arrows, labels and XML agree with the forward computation."},{"id":"C5","verdict":"supported","reason":"Column normalization, magnitude initialization, directional adaptation and materialization match Sections 3.2 and 4.1 of [DoRA v3](https://arxiv.org/html/2402.09353v3)."},{"id":"C6","verdict":"supported","reason":"Norms, reconstructed vectors and column-wise parameter count are correct. The helper explicitly rejects zero columns."},{"id":"C7","verdict":"supported","reason":"Frozen four-bit adaptation, NF4 construction and the attributed 65B/48GB result match the abstract, Section 3 and appendices of [QLoRA](https://arxiv.org/pdf/2305.14314)."},{"id":"C8","verdict":"supported","reason":"Scale precisions, block sizes and optimizer paging match Section 3 of [QLoRA](https://arxiv.org/pdf/2305.14314); capacity and transfer limitations are appropriately bounded."},{"id":"C9","verdict":"supported","reason":"The live [integration article](https://huggingface.co/blog/4bit-transformers-bitsandbytes) and matching saved revision explicitly distinguish storage from BF16 computation and adapter-only updates."},{"id":"C10","verdict":"supported","reason":"Independent arithmetic confirms 4.126953125 bits per weight and 8,452 bytes. Omitted offsets, metadata and other allocations are explicitly disclosed."},{"id":"C11","verdict":"supported","reason":"Distributivity preserves the reconstructed-base projection; mapping an arbitrary merged matrix onto a finite quantization representation generally changes it."},{"id":"C12","verdict":"supported","reason":"Vector-wise INT8, FP16 outlier contributions, threshold 6.0 and approximate prevalence match Sections 3.1–3.3 of [LLM.int8() v2](https://arxiv.org/html/2208.07339v2)."},{"id":"C13","verdict":"supported","reason":"The reconstruction objective, damped inverse Hessian, compensation and Cholesky/block formulation match Sections 3–4 and Algorithm 1 of [GPTQ](https://arxiv.org/pdf/2210.17323)."},{"id":"C14","verdict":"supported","reason":"The distinction between saliency experiments and activation-informed scaling, including the search procedure, matches Sections 3.1–3.2 of [AWQ v5](https://arxiv.org/html/2306.00978v5)."},{"id":"C15","verdict":"supported","reason":"Both dot products and the reciprocal diagonal-scaling identity are correct. Quantization need not preserve that identity; the example makes no algorithm-performance claim."},{"id":"C16","verdict":"supported","reason":"Complete code inspection supports the deterministic inputs, standard-library scope, assertions, tie handling and packing description. Supplied isolated execution is consistent with these limited claims."},{"id":"C17","verdict":"supported","reason":"Direct multiplication confirms the adapter contribution, merged matrix and output [5, 9.5]; the supplied execution agrees."},{"id":"C18","verdict":"supported","reason":"All reconstruction constants correctly round [QLoRA Appendix E](https://arxiv.org/pdf/2305.14314). Code implements the disclosed teaching approximation and keeps scales/counts separate from packed indices."},{"id":"C19","verdict":"supported","reason":"The complete package remains within mechanism explanation and synthetic arithmetic; neither the diagram nor lab establishes training quality, memory fit or deployment benchmarks."},{"id":"C20","verdict":"supported","reason":"Source versions and dates match reopened records. ResearchNode returned the exact cited public document/revision. The ledger distinguishes source passages, derivations and code inspection."}],"package_sha256":"1959ea21a3134c8c1ba910fd8e03c37ef6d2450ba0127e3a1ff5d69e189ac134","reviewed_at":"2026-11-16T18:37:46.553421+01:00"},"visual":{"verdict":"pass","summary":"Reviewed all 79 rendered screenshots in the supplied attempt-2 manifest against the five style references and supplied content, including every desktop section, both mobile widths, scroll-panel positions, question states and nine print pages. No material visual issues found. Dark styling, syntax colours, rounded question badges, numbered references with full URLs and spacing are consistent. Diagram labels, arrows and supplied text alternatives agree. Mobile overflow stays within panels, with content reachable in the scrolled views. Print content is readable and complete. Recorded browser checks cover keyboard interaction, copying, contents links and downloads; these results were supplied, not independently rerun. This review does not certify factual accuracy.","issues":[],"claims":[],"package_sha256":"1959ea21a3134c8c1ba910fd8e03c37ef6d2450ba0127e3a1ff5d69e189ac134","reviewed_at":"2026-11-16T18:37:46.553438+01:00"}},"checks":{"package_sha256":"1959ea21a3134c8c1ba910fd8e03c37ef6d2450ba0127e3a1ff5d69e189ac134","passed":true,"widths":[1440,390,320],"code_checked":true,"code":[{"language":"python","syntax":"passed","sha256":"3c0ed56ce8e99001f110e35b319e07d78c4eec0e6f4379ef393c81d6636c147a"}],"execution":[{"file":"lab.py","sha256":"8252fe461807b5a9c12ec73919bd7db93501655454236e1ddd3c6838abc2f875","returncode":0,"output":"selected projection counts: 67108864 524288\nsynthetic merged output: [5.0, 9.5]\nsynthetic block size, RMSE, index bytes, scale count: 4 0.027085550417970918 5 3\nsynthetic block size, RMSE, index bytes, scale count: 8 0.027085550417970918 5 2\nideal bits per weight with nested scales: 4.126953125\n","environment":"CPython 3.12 isolated container; no network, no host mounts; 128 MiB; 1 CPU"}],"drawio":["adapter-branches.drawio"],"print":"rendered"},"published_at":"2026-11-16T17:37:50.528Z"}