What a cited model version does not record: a two-round correspondence audit of four commercial language-model vendors

Supplementary evidence for the article "What a cited model version does not record: a two-round correspondence audit of four commercial language-model vendors". The study audits the correspondence between what four commercial vendors declare in their own public documentation and interfaces, and what their products deliver when the declared object is exercised. The vendors are Anthropic, DeepSeek, Google and OpenAI. Four strata of evidence were collected for each product: vendor documentation, the conversational interface, the account data export and the programmatic interface. Two collection rounds were conducted, from 23 to 25 July 2026 and from 18 to 19 August 2026, with a fixed-files complement on 26 July 2026. The first round was exploratory. The second was executed prospectively against a protocol frozen and sealed before collection with the SHA-256 digest 4ec3b683f2f851e062aab8e53521ce72c4375428872ab156dd2396afa0edd35c, which is shipped here so that any reader can recompute it. This deposit contains the consolidated evidence register, 164 numbered observations, 108 from the first round and 56 from the second. Each line carries the requirement tested, the class of the observation, the claim in prose, a literal quotation of the declaration or of the observed output, the address, the access date and time zone, the local artefact and its SHA-256 digest, the public archive addresses where applicable, and a verification state. Conformity is recorded on the same terms as divergence: 33 lines record conformity, a declared baseline or the absence of a feature. The raw evidence package contains the captured vendor documentation as single-file MHTML archives and as PDF, the programmatic payloads with full response bodies and response headers, screen captures and verbatim transcriptions, the frozen protocol with its pre-registration record, the collection scripts, and per-file manifests with SHA-256 digests. The Supplementary Information, in fourteen notes, documents the standardised stimulus verbatim, the requirements tested, the classification of observations, the evidence and archiving discipline, the instrument control used to test declared non-persistence, and the data availability with its known gaps. Account data export packages produced by the four vendors are not published, because they contain personal data of the first author. For each of them the file name, size, SHA-256 digest and a structural description are provided. Extracts limited to a single observation are supplied on request by the corresponding author. Redactions applied to published artefacts before deposit are listed, with the digest of each unmodified original, in the sanitisation record. Rights are set out in LICENSE.txt. Material authored by the authors is licensed CC BY 4.0. Material captured from the four vendors, that is saved documentation pages, screen captures of product interfaces and programmatic response bodies and headers, is not covered by that licence: rights remain with their respective owners, and the material is reproduced only in the quantity needed to verify a specific claim of the article. Known gaps are declared in README.txt. Public archiving applies only to vendor pages reachable without an account; observations made in authenticated interfaces or against programmatic endpoints are not archivable by a public service, and for those lines the evidence is the hashed screen capture or the saved response body. Twenty-two lines are reported as not independently verifiable from this package, and the reason is stated for each. Three shipped documents carry an earlier working title of the manuscript and are not edited, because one of them is the frozen protocol whose digest is the pre-registration. Cite the article. This package is its supplementary evidence.