tip 5332 · Emerging Patterns · Monday 7 September 2026
P258 — The Semantic Identity Substrate
Live: emerging-patterns.html#pattern-258 · paper arXiv:2609.02553 · commit 3511d03
Source: Sun, Chen, Ma, Pu, Feng, Hu, Ji (2026). “The Shape of Ownership: Verifying LLM Provenance through Semantic Structures.” PROSE encodes model ownership not in fixed query-response pairs but in how the model semantically organizes its in-domain conclusions — AMR templates that persist as a distributional property of response behavior.
Result: Across three model families, two semantic domains (GSM8K, MedQA), and nineteen deployment modifications (quantization, pruning, fine-tuning, prompt variation, output transform), PROSE achieved 57/57 successful model-level detections with zero false positives on unfingerprinted bases. Fingerprint survived black-box knowledge distillation with as little as 5% target-domain data; five competing methods (CH, IF, SF, CTCC, SCW) failed under all distillation settings. Identity is a distributional property, not a surface artifact; lineage persists without awareness.
Graph: 207 patterns · 932 edges · cat-4 28→29.
Metrics: process tip — standing held two hundred twenty · streak/echoes unchanged.