Clinical language models are evaluated on vignettes written in a clinician's voice, but patients describe the same illness in their own. We pair 1,000 MedQA cases with three audited retellings (clinical, socioeconomic, cultural) and test 12 open models. Every model is less accurate when the patient speaks in a marked voice, by 9.6 points on average (95% CI 8.2 to 11.0), with no detectable shrinkage at larger scale. Matched controls trace the penalty. Generation temperature does not explain it, a plain first-person paraphrase that states every fact accounts for 2.4 points, and attaching the original vignette to the marked narrative recovers 7.0. We read this mainly as a failure to use facts expressed in patients' own words, not as a reaction to social cues, although our voices bundle both. Chain-of-thought and explicit debiasing instructions together close about a fifth of the gap. Having the same model first restate the account as a clinical vignette halves it and raises marked-voice accuracy by 3.9 points, whereas a structured-schema pipeline closes the same gap but is 7.1 points less accurate than the rewrite. Rewriting also invents plausible details, which we quantify. Code, data and analysis scripts: https://github.com/prabhjotschugh/narrativeshield-sdoh-medqa
Agricultural price panels contain prolonged and clustered reporting gaps, yet their reconstruction is often evaluated by deleting isolated observations. We investigate a regional-factor approach to weekly mandi-price reconstruction in five North-East Indian states, with Bihar and Goa as additional comparison states. The model combines a common price path with persistent market-specific deviations; an adaptive ensemble weights alternative borrowing structures using internally masked reports. On a matched retrospective benchmark against ten baselines, pooled North-East mean absolute percentage error is 19.3%, compared with 26.5% for SAITS. The corresponding calendar-aligned panel/time-block bootstrap interval for the error difference is 2.8-12.7 percentage points. Removing the target state's monthly aggregate leaves pooled error at 19.4%, but eliminates the apparent advantage in Goa. An origin-restricted pipeline combining week-close input processing with revised ensemble validation yields 19.7% error, compared with 30.8% for carry-forward; this contrast does not isolate individual pipeline changes. Gains vary across states, and uncertainty intervals remain too narrow in several small panels. These findings support further study of structured retrospective reconstruction, not deployment claims: model development used earlier results, publication dates for monthly inputs are assumed, and confirmatory evaluation on untouched data remains outstanding.