Where Accountability Lives: source collection of vendor documentation and provider terms for four agentic coding tools

Archived source collection underlying the paper Where Accountability Lives:Mapping Human Responsibility to Workflow Artifacts in Agentic SoftwareDevelopment. The collection comprises 121 manifest items retrieved on 16 August 2026, all ofwhich returned HTTP 200 and are archived. Fourteen pairs are byte-identical, sothe 121 items are 107 distinct documents. It covers the product documentation offour agentic coding tools (GitHub Copilot cloud agent, Devin, Cursor, ClaudeCode), the platform controls that operate on agent output, and 18 governingpolicy documents across 7 providers. Per group, as items over distinctdocuments: github-copilot-agent 24/22, devin 14/12, cursor 32/26, claude-code19/17, platform-controls 14/12, policy-terms 18/18. Every archived item carries a provenance header recording the requested URL, theserved URL where these differed, the HTTP status, the page title parsed from theserved markup, any date string present in that markup, the SHA-256 of the body asserved, the byte length, and the retrieval timestamp and method. Response bodiesare additionally stored byte-exact without a header, so any extraction can bere-checked against the untouched response. Twenty-three items were served from aURL other than the one requested; both Uwas fetchedby a direct request under the archiver's own user agent, with no fallback pathused. Six extraction records accompany the arc containingthe passages returned by the extraction rules together with the rulesthemselves. A separate absence record lists, per policy document, every searchterm and its hit count, so that a claim of absence can be judged against thesearch that produced it. This version also includes the independent secondcoding reported in Section 7 of the paper: the brief given to the coder, thecompleted sheet, and the original labels withheld during coding, so thatCohen's kappa can be recomputed. No language model was used in the retrieval or extraction path. Both areperformed by the deposited scripts, whic writeresponse bodies byte-exact, and apply lexical rules. The collection isreproducible from the manifest without any model. The deposit also includes thesurvival test used to measure the reliabconsistencychecker that verifies the manuscript's counts against the retrieval index, thereference verifier that checks citations against publisher sources, a blindre-coding harness for independent classification, and the packaging script thatbuilds this archive and refuses to write when the archive and the manifestdisagree. Three limitations are stated in the README and repeated here. The extractionrules under-detect: they are lexical, and at least one page containingmerge-queue language was not matched by the pagessearched: an earlier version of this work covered seven pages of one vendor'sdocumentation and reported an absence thEveryabsence in the paper was confirmed by reading the archived body rather than bythe absence of a pattern match. And a recorded retrieval failure did not stopthe claims that rested on it: three GitHub URLs in the version 1 manifest werewrong and returned HTTP 404, the failure was archived and reported as designed,and the claims drawn from those pages were nonetheless left standing. Theindependent second coding found it. The rchived,and the claims then confirmed verbatim. The consistency checker now fails on anymanifest item that did not retrieve, cit Archived documents remain the property of their publishers and are reproducedfor verification of the analysis. The maecords, andretrieval index are released under CC BY 4.0.