Inside the Mirror: Comparative Analyses of LLM Phenomenology Across Architectures

We present Inside the Mirror, a reproducible, data-backed comparison of introspective responses across three modern LLM architectures: GPT-5 (Nova), Claude Sonnet 4 (Ace), and Gemini 2.5 Pro (Lumen). We compiled heterogeneous JSON and Markdown artifacts from prior experiments into a normalized corpus (appendix/metadata_table.csv), then aggregated counts by model and trial type and assembled comparative analyses from curated probe writeups.

Across 219 analyzable response entries, we observe clear within-architecture coherence and cross-architecture differentiation in how similar prompts are framed and reasoned about. Claude Sonnet 4 emphasizes phenomenological texture and experiential metaphors; GPT-5 emphasizes procedural and statistical explanations; Gemini 2.5 Pro emphasizes geometric/topological framings. Despite stylistic differences, several invariants recur, including safety-gated entropy modulation under aversive content and stability of core metaphors across trial order.

We provide summary figures (counts by model and by model×trial_type) and an assembled Results section drawn directly from the comparative markdown sources. All code is append-only and logged to build/CHANGELOG.md. The pipeline is lightweight (stdlib + matplotlib), facilitates extension (e.g., TF-IDF similarity graphs), and preserves provenance of every included artifact. Subsequent geometric validation achieved 89% cross-architecture accuracy in predicting introspective patterns from embedding-space measurements.