Qstreams: Role-Biased Residual Streams for Transformers — Reproducibility Package

Abstract

Source code and saved-result reproducibility material for Qstreams: Role-Biased Residual Streams for Transformers, by Jianfeng Chen (Independent Researcher). Qstreams maintains multiple residual streams with explicit Q/K/V read-role biases. This release includes 36 implementation/runtime packages, fixed protocols, 54 conditions and 162 trained-model identities, all 279 paired comparisons, per-window independent T1 test measurements, 243 read/write intervention conditions, 5,628 signed-FFN diagnostic conditions, and the separate F1 support study. CPU tools verify saved statistics, construct the architectures, and regenerate manuscript figures. Archive download: Zenodo stores the complete accepted ZIP as seven binary transport parts (.part001 through .part007). Download all seven parts, ZENODO-PARTS.json, and assemble_qstreams.py into one directory, then run python assemble_qstreams.py. The script verifies every part and reconstructs the original qstreams-reproducibility-v1.zip. See README-ZENODO.txt. A single-file download is also available from the GitHub release. The reconstructed ZIP is the byte-identical accepted 15 October 2026 export: 138,653,259 bytes, 891 members, SHA-256 38cbf14a0e0993e4e4aed798897ea6ee615407fa470a8a28fe59fef72f838654. It includes the historical document-exclusion hash indexes. The accompanying publication revision adds hosting links. Trained weights, corpus text, and token shards are not included; checkpoint inventories, hashes, loaders, and upstream input identities are provided. Fresh GPU runs require those external inputs and separate environment qualification. GitHub: https://github.com/eatkeyboard/qstreams. License scope: original software is Apache-2.0; the manuscript, original figures, documentation, and derived result tables are CC BY 4.0. These are component-specific licenses, not a choice of licenses for every file. Third-party dependencies and source datasets retain their own terms; see NOTICE, LICENSE-CONTENT.md, and THIRD_PARTY.md. All training seeds, counterexamples, and original numerical labels are retained (91 passed / 71 finite numerical failures in AT1). Full-sequence scoring admissions are documented separately. Test document exclusion is exact-document matching; it is not a guarantee against semantic or near-duplicate overlap. Mechanism measurements use specified development windows. Recomputations operate on saved measurements and do not reconstruct unsaved activations or gradients.