ARGIRA v2 Lexica Subcorpus: Visual Style Annotation and Heuristic Characterization — Evidence Package (v1.1.0)

Evidence package documenting the LEXICA experimental sequence, from A to C.2.This package documents the LEXICA experimental sequence within the ARGIRA project. ARGIRA is a deterministic visual-characterization heuristic designed to study calibrative transparency in automatic alt-text systems. The study uses a 23-image exploratory subcorpus from the Lexica source. LEXICA-A.Human annotation of visual style.A closed taxonomy was defined before running ARGIRA. The categories are: photorealistic, painterly, illustration flat, and digital composite. The annotations were frozen before any heuristic analysis. LEXICA-B.Unmodified execution of the ARGIRA heuristic.This phase provides a descriptive comparison between human-assigned visual style and ARGIRA outputs, including predictions, confidence values, and margins. LEXICA-C.1.Exploratory analysis of disagreements between human descriptions and heuristic outputs.These differences are treated as two different representations of visual structure, not as classification errors. LEXICA-C.2.Added in version 1.1.This phase projects the 23 Lexica images onto a decision frontier created in an earlier independent experiment, called Phase A. Phase A compared 107 human paintings with 20 earlier Lexica images from a visually homogeneous painterly collection.The frontier was not retrained, modified, or validated using the 23 images in this package. It was only used as a projection reference. The result suggests that the previous separation does not generalize as a detector of image provenance, meaning human versus AI origin. Instead, it appears related to the visual characteristics of the original painterly collection. A post-hoc coherence audit between LEXICA-C.1 and LEXICA-C.2 is included. This audit does not modify any previous result. This is an exploratory evidence package. It is not a trained classifier, not a benchmark dataset for AI detection, and not a statistical validation of ARGIRA performance. The subgroup sizes are small. Results should be interpreted as a characterization of heuristic behavior, not as estimates of general performance. Version 1.1.0.Adds LEXICA-C.2 and the Section 8 coherence audit.No file content from version 1.0.0 was modified.