Documentary Provenance in the CFPB Consumer Complaint Database: A Working Corpus and Exploratory Audit (2011–2026)

This deposit documents an exploratory, independent research line examining documentary provenance -- whether a record can be reliably attributed to the decision event and party it claims to represent -- using the CFPB Consumer Complaint Database (2011-2026) as an empirical site. It is distinct from documentary completeness (whether a record contains enough content fields to support later account-giving), the construct examined in the author's prior ATRS/Dutch-register audit work. Source data: official CFPB bulk download (files.consumerfinance.gov/ccdb/complaints.csv.zip), N=17,004,291 records. Working corpus: 2,040,778 records with both a consumer narrative and a company response present. Key findings (exploratory, not causal): Exact-duplicate consumer narratives: 42.73% of the working corpus (full population). - Near-duplicate narratives (MinHash LSH, reproducible sample n=60,000, seed=42): 38.24%. - The institution-side "Company public response" field is a fixed 11-category pick-list, not free text; 91.98% of records select "chooses not to provide a public response" -- a completeness/visibility finding, not a provenance one. - Decision-level test: duplicated (templated) narratives are associated with markedly lower odds of a monetary-relief outcome than unique narratives. Uncontrolled odds ratio 0.0051 (~197x); after controlling for product category (credit reporting vs. other products, Cochran-Mantel-Haenszel), common odds ratio 0.0227 (~44x). The association survives control but is not established as causal. This deposit includes the analysis scripts, run logs, a corpus codebook, and a working paper draft, for independent reproduction. Due to file-size constraints on this upload, the large source/derived data files (the ~1.3GB official CFPB bulk ZIP and the ~2.7GB working-corpus CSV) are not included here; they are fully reproducible from the cited public CFPB source using the included scripts. All limitations, including unaddressed confounds, are stated explicitly in the included run logs and working paper -- this is early-stage, exploratory work, not a validated causal analysis.