Auditable RAG Credit Signals from Auto-Retail Filings: A Validation Framework for Dealer Risk Models

Captive auto-finance companies extend wholesale (floorplan) credit to thousands of franchised dealers whose financial condition is reported largely through unstructured narrative. Traditional dealer credit models rely on structured financial ratios and behavioural data, often discarding the forward-looking "soft" information that management discloses in text. We propose a retrieval-augmented generation (RAG) framework that converts public
auto-retail filings into grounded, point-in-time credit features, including covenant-headroom tone, floorplan-risk language, liquidity adequacy, and going-concern flags. Each non-null feature carries a verbatim citation to its source and passes a deterministic quote-verification gate, making the extractor a validation object in its own right rather than a free-form language model. We evaluate the extractor as a feature-generation component for credit risk models: the benchmark is a traditional structured-ratio model, the challengers add RAG features or non-LLM textual baselines from the same retrieved passages, and all comparisons use out-of-time validation, correlated-AUC testing, issuer-block bootstrap inference, split and label-threshold sensitivity, feature attribution, and model-governance diagnostics. On 369 issuer-quarter observations for ten US-listed auto retailers, grounded RAG features add a small amount of discrimination under a fixed gradient-boosted learner: out-of-time AUC rises from 0.852 to 0.857 and remains positive after adding an equity-volatility market control (0.845 to 0.856). Generic non-LLM text baselines do not improve the structured model in this public panel: keyword/tone features reduce AUC to 0.822 and TF-IDF/SVD features to 0.837. The RAG lift is modest, statistically fragile, and not a final dealer-default performance estimate, but the exercise demonstrates an auditable extraction and validation design that can be deployed on a captive's internal dealer financials with realised default flags.