Get started
Work

Code and run records for: Target Headroom, Not Semantics, Governs a Domain-Gap Ablation

Year: 2026
Type: software
Abstract: Code and complete run records for the study "Target Headroom, Not Semantics, Governs a Domain-Gap Ablation". The study shows that a destructive-control decomposition of a domain gap returns a value bounded by the target model's clean performance, so where that performance is low the reported number describes the source model and the corruption strength rather than the gap between domains. Across 1,992 attributed conditions, control identity and severity account for 0.341 of the variance in the reported share, target headroom raises that to 0.720, and the identity of the domain pair adds 0.011 beyond it. The deposit contains the experiment driver, the control and attribution code, the analysis scripts, and one JSON run record per benchmark family per seed for digits, natural images and PACS, over three seeds. Each record carries the four raw measurements from which a share is computed, so any derived quantity can be recomputed independently of this code. verify_paper_numbers.py reproduces every number reported in the paper, in the order the paper presents them, and needs no GPU, no dataset download and no retraining. Reproducing the runs from scratch does require the public benchmark datasets and roughly an hour of CPU per seed per family. check.py applies the same diagnostic to an ablation that has nothing to do with this paper. Given the four clean and control measurements a reader already has, it returns the share, the guaranteed lower bound on source damage, and whether the bound is vacuous: python check.py --ms-clean 0.85 --mt-clean 0.32 --ms-ctrl 0.71 --mt-ctrl 0.29 The associated manuscript has been submitted to a peer-reviewed journal. Publication details will be added here once they exist. The code is released under the MIT License. The run records in runs/ may be reused under CC BY 4.0 with attribution.
DOI: 10.5281/zenodo.21900329
Language: English
Open Access status: green
OpenAlex ID: W7202267326
Cites:
Cited by:
Related to: 10