The article introduces a compact, locality-sensitive fingerprinting method for AI agent skills that uses a multi-bank SimHash to generate fixed 120-byte signatures. Instead of relying on cryptographic hashing which is sensitive to minor edits, this approach projects each component of a skill—prompt, code, and tools—to bits for comparison via Hamming distance.
- The fingerprint maintains a per-component triple structure rather than a single score, allowing it to recover skill-family identity through paraphrase, renaming, and refactoring while localizing which component carries the reuse.
- It achieves an area under the ROC curve (AUC) of 0.974 over 4,950 pairwise comparisons while using 77x fewer bits than the embedding it approximates.
- On a 906-skill injection benchmark, the fingerprint recognizes injected skills as tampered copies of a known base and localizes the change.
The authors argue that this method supplies lineage and structural identity for a skill registry, serving as an identity signal complementary to behavioral verification rather than a safety verdict.