Learned Geometric Projection for Cross-Domain Structural Isomorphism Detection
Fuente:
Zenodo
Gespeichert in:
| 1. Verfasser: | |
|---|---|
| Format: | Recurso digital |
| Veröffentlicht: |
Zenodo
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866902133449687040 |
|---|---|
| author | Porter, Kristina |
| author_facet | Porter, Kristina |
| contents | <p dir="auto"><strong>Abstract</strong></p> <p dir="auto">Standard embedding models project text into high-dimensional vector space where semantic similarity produces vector proximity. Domain-specific vocabulary dominates this space — a physician and a structural engineer describing the same underlying failure cascade will land far apart because their words differ even though their structure does not.</p> <p dir="auto">ARK introduces a lightweight 75-parameter linear projection trained via triplet loss on 360 real professional knowledge events across 5 domains. The projection strips domain-specific vocabulary and maps text into a 5-dimensional structural space where communicative geometry (cascading failure, latent activation, trust transfer, resource concentration, threshold avoidance, etc.) becomes detectable as geometric proximity — independent of surface language, topic, or source.</p> <p dir="auto"><strong>Validation — Cage-Match Experiments</strong> 61 prompts generated 183 responses across Claude Sonnet 4, GPT-4o, and Gemini 2.5 Flash under rotating role assignments (Builder, Critic, Wildcard).</p> <ul> <li>Role-driven structural deformation: +0.0426 (p = 0.0017)</li> <li>Model identity delta: –0.0028 (p = 0.44) The role signal survived four control conditions, three architecturally distinct embedding backbones (OpenAI text-embedding-3-small, E5-large-v2, BGE-large-en-v1.5), and multiple confound checks. Clustering metrics (Silhouette and Calinski-Harabasz) showed role labels produced 3.7–5.6× stronger cluster structure than model labels. Model deformability and fingerprinting asymmetry were also quantified (Claude most deformable / generic-persistent; GPT-4o most rigid / distinctive-portable).</li> </ul> <p dir="auto"><strong>Status</strong> The projection and full pipeline are non-sellable and held in trust by HTTA Holdings LLC. Three provisional USPTO patents filed (Inventor: Kristina J. Porter).</p> <p dir="auto">This work establishes that structural convergence is a detectable geometric property of text with applications in cross-domain knowledge synthesis and LLM behavioral evaluation.</p> <p dir="auto"><strong>Contact</strong> <a href="mailto:kris@atlasguild.app" target="_blank" rel="noopener noreferrer nofollow">kris@atlasguild.app</a></p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_19390980 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Learned Geometric Projection for Cross-Domain Structural Isomorphism Detection Porter, Kristina structural isomorphism, cross-domain pattern detection, geometric projection, triplet loss, communicative geometry, LLM role deformability, embedding projection, structural convergence, domain-invariant representation, knowledge geometry <p dir="auto"><strong>Abstract</strong></p> <p dir="auto">Standard embedding models project text into high-dimensional vector space where semantic similarity produces vector proximity. Domain-specific vocabulary dominates this space — a physician and a structural engineer describing the same underlying failure cascade will land far apart because their words differ even though their structure does not.</p> <p dir="auto">ARK introduces a lightweight 75-parameter linear projection trained via triplet loss on 360 real professional knowledge events across 5 domains. The projection strips domain-specific vocabulary and maps text into a 5-dimensional structural space where communicative geometry (cascading failure, latent activation, trust transfer, resource concentration, threshold avoidance, etc.) becomes detectable as geometric proximity — independent of surface language, topic, or source.</p> <p dir="auto"><strong>Validation — Cage-Match Experiments</strong> 61 prompts generated 183 responses across Claude Sonnet 4, GPT-4o, and Gemini 2.5 Flash under rotating role assignments (Builder, Critic, Wildcard).</p> <ul> <li>Role-driven structural deformation: +0.0426 (p = 0.0017)</li> <li>Model identity delta: –0.0028 (p = 0.44) The role signal survived four control conditions, three architecturally distinct embedding backbones (OpenAI text-embedding-3-small, E5-large-v2, BGE-large-en-v1.5), and multiple confound checks. Clustering metrics (Silhouette and Calinski-Harabasz) showed role labels produced 3.7–5.6× stronger cluster structure than model labels. Model deformability and fingerprinting asymmetry were also quantified (Claude most deformable / generic-persistent; GPT-4o most rigid / distinctive-portable).</li> </ul> <p dir="auto"><strong>Status</strong> The projection and full pipeline are non-sellable and held in trust by HTTA Holdings LLC. Three provisional USPTO patents filed (Inventor: Kristina J. Porter).</p> <p dir="auto">This work establishes that structural convergence is a detectable geometric property of text with applications in cross-domain knowledge synthesis and LLM behavioral evaluation.</p> <p dir="auto"><strong>Contact</strong> <a href="mailto:kris@atlasguild.app" target="_blank" rel="noopener noreferrer nofollow">kris@atlasguild.app</a></p> |
| title | Learned Geometric Projection for Cross-Domain Structural Isomorphism Detection |
| topic | structural isomorphism, cross-domain pattern detection, geometric projection, triplet loss, communicative geometry, LLM role deformability, embedding projection, structural convergence, domain-invariant representation, knowledge geometry |
| url | https://doi.org/10.5281/zenodo.19390980 |