The XZ Structural Honesty Audit Corpus: 832 Findings from 31 Rounds of Multi-Model Verification

Structured dataset of 832 findings (812 unique finding IDs) extracted from 199 seat reports across 31 audit rounds of the XZ Protocol (v0.13 through v0.74), produced by independent AI systems from five vendor families (Anthropic, OpenAI, Google, Moonshot, Zhipu). The primary empirical series -- 25 consecutive rounds (R7--R31) with Class A (byte-false claim) defects present -- yields 102 extracted Class A findings across 19 of those 25 rounds. The corpus contains 136 Class A findings in total. The construction procedure (extract_corpus.py v0.6, an 18-pattern regex extractor with 52 manual verdict overrides) ships with the deposit as a reproducible pipeline. The deposit includes: the dataset card (paper v1.3), four structured data files (findings.csv, reports.csv, coverage.csv, summary.json), the extractor script, and a SHA-256 manifest. Four audit rounds (R1--R4) were conducted across three non-producer model families (Kimi K3, GPT-5.6 Sol, Gemini 3.1 Pro); the carrying seat returned CONDITIONAL PASS at R4. This paper makes no effectiveness claim about any component of the Structural Honesty program. Prepared with Claude Opus 4.6 (Anthropic) as analytical and drafting instrument under the author's direction and final authority.