Cross-Architecture Per-Layer Residual Stream Alignment via Tier-B Fingerprints: 84 Model-Pair Matrices Across 9 Open LLMs Show a Universal Middle-Depth Slot

I observe nine open large language models (Qwen2.5 3B/7B/14B, Falcon3 7B, Granite 3.1 8B, OLMo2 7B, Gemma 2 9B, Mistral 7B v0.3, Yi-1.5 9B, DeepSeek-R1 7B) at every transformer block via HuggingFace output_hidden_states, build a per-layer residual stream fingerprint for each block on ten multilingual prompts, and compute pairwise cosine similarity across all 84 ordered model-pair matrices.

Cross-model alignment is not uniform along network position. Among the 84 matrices, 54 contain their top matched block in the 50 to 60 percent depth slot (6.4x the uniform-null expectation), against 18 in the 0 to 10 percent slot. The single top cross-architecture pair is Qwen2.5-7B block 15 against Falcon3-7B block 16, cosine similarity 0.868.

A second observation pass at the per-head attention level (Tier-C) on the same nine networks replicates the depth concentration independently: the top one percent closest 5D head fingerprints out of 74,112 head-to-head pairs concentrate at the same 50 to 60 percent slot (139 of 754 top-percentile matches).

The middle-depth slot is therefore an architecture-invariant target for distillation alignment, model merging, and per-layer interpretability probes.

Companion records: Mercury method (10.5281/zenodo.20313154), Mercury Viewer (10.5281/zenodo.20313150), Discovery (10.5281/zenodo.20313748), Scaling (10.5281/zenodo.20325676), Qwen family cross-model dim (10.5281/zenodo.20346626), Mercury MCP v0.1 (10.5281/zenodo.20352085).

Open data and code: https://github.com/norika1207-lab/mercury-paper-handoff