A Target-Blind Framework for Corpus-Based Cognitive Profiling: Leakage Control, Benchmark Replication, and Prevalidation Toward External Calibration

Naturalistic language and problem-solving traces may contain information about cognitive processes, but attempts to map such traces onto psychometric constructs are especially vulnerable to target leakage, post-hoc feature construction, source dependence, and overinterpretation of correlation as calibration. This methods/prevalidation study describes AERIS EXTCAL, the external-calibration workstream of AERIS, a project-specific modular architecture for discovery, structured reasoning, evidence handling, provenance, validation, and fail-closed workflow control in AI-assisted analytical work. EXTCAL is a target-blind framework for corpus-based cognitive profiling designed to separate feature construction, benchmark replication, development, quasi-external evaluation, and strict external validation. Two public-source case studies were used to test the framework without fitting a global intelligence mapping. In a German Map Task corpus (D043; 28 adults, 14 dyads), participant-level auxiliary naturalistic vectors were materialized while subject/target files remained semantically sealed; dyad orientation, speaker-role context, source-specific exclusions, and dyad-only floor-transfer measures were enforced before feature freeze. In a separate lexical-proxy dataset (D016), the published Study-2 development/test structure was rebound as a method benchmark (53/53), and the historical proxy behavior was reproduced: test correlations with measured IQ were r = .702 for WAT, r = .738 for SpanLex, r = .700 for SCIRT, and r = .798 for the median estimator; the median estimator yielded MAE = 7.72 and RMSE = 9.44 IQ-score units in the published benchmark scale. Critically, D016 was not repurposed to tune AERIS, and D043 targets remained sealed. The resulting negative gate finding was that the frozen V1 development scope contained no fit-eligible same-person pairing of a multi-domain AERIS naturalistic profile with an independent standardized IQ target. Thus, the engineering and reproducibility framework is operational, but individual standardized-IQ predictive validity is not established. A low-dependency open-criterion V2B protocol and a future standardized-target V2A pathway are specified for empirical validation.