AS-DCCF: raw per-seed run outputs

Raw per-seed training outputs for the AS-DCCF study. This record holds the unaggregated run artifacts that every number in the accompanying paper was computed from. It is the evidence layer: each file is a single training run's own output, not a summary. Contents. 6,624 files, 1.56 GB compressed, covering the study's measurement tracks: protocol_v2 (confirmatory family), protocol_v3 and protocol_v3_1 (cost, fidelity, adaptive-controller and transfer tracks), protocol_v3_2 (the same-batch remediation batch), protocol_v3_3 (the XSimGCL baseline promotion), benchmark_expansion, reference_final, and the audit tracks. Each run contributes: *.csv — per-epoch metrics, with the run's complete resolved configuration embedded in every row *.json — run summary including the hardware fingerprint and test-evaluation count *.timing.json — timing sidecar under a single declared caliber *.fidelity.json — four-arm probe records, for probe runs only Benchmarks. Amazon, Tmall and Gowalla implicit-feedback splits, all under one frozen training budget. Hardware. Two GPU classes, kept separate by the study's fairness contract: an NVIDIA A800 80GB PCIe for the confirmatory efficiency family, and NVIDIA RTX 4090 for the v3 cost tracks and both follow-up batches. No ratio is formed across them. This record is not self-contained. The code, configuration locks, preregistered protocols, prepared data splits and analysis summaries live in the companion record https://doi.org/10.5281/zenodo.21989855. Download both and unpack them over one another to reconstruct the tree the paper was built from. Integrity. as-dccf-raw-runs.tar.gz, SHA-256 ffd7b07846bc7e91d61e743efa8d2d9288d21147aa2998387d7b7d5a989ba670 Corresponding author: Wei Li ([e-mail]).