NervaptaCONFIDENTIALNervaptaCONFIDENTIALWe train transformers on the Lorentz hyperboloid. Knowledge self-organizes into navigable coordinates. You can query it. You can erase it. Surgically. In milliseconds.
| Capability | SQL / RDBMS | Vector Database | Nervapta |
|---|---|---|---|
| Knowledge storageWhere facts live | Structured tables, explicit schema | Embedding vectors, approximate | Geometric coordinates on hyperboloid NATIVE |
| Exact lookupFind a specific fact | Primary key / indexed column | Not supported — approximate only | Navigate to sector + x₀ coordinate GEOMETRIC |
| Semantic searchFind related concepts | Not supported natively | Cosine similarity, ANN index | Angular proximity within sector EXACT DIST. |
| Language generationProduce coherent text | Not applicable | Retrieval only, no generation | 41.08 PPL @ 7B — beats Euclidean +3.56 PPL |
| Knowledge deletionGDPR / compliance erasure | DELETE row — does not erase model memory | Remove embedding — model unchanged | Zero tokens via cone deletion · 6.9ms · MI receipt 8,106× SPECIFIC |
| Erasure verificationProve what was deleted | Audit log (row-level) | No — embeddings cannot be audited | Cryptographic MI receipt · 0/13 recall probes GDPR RECEIPT |
| Entity-level separationDistinguish individuals | Explicit foreign key | Approximate — similar names cluster | 136/136 entity pairs separable · cosine dist > 0.94 100% RECALL |
| ETL / indexing requiredPipeline overhead | Schema design + data ingestion pipeline | Embed → index → sync on every update | Zero — forward pass IS the index 0 PIPELINES |
| Scaling lawBehavior at larger scale | Scales with storage, not intelligence | ANN accuracy degrades at high dimensions | x₀ ∝ N^1.00, R²=0.968 · confirmed 200M → 7B → 28B LINEAR LAW |
| Coordinate queryabilityQuery by address, not content | Row/column address | No addressing — vector-space only | Sector + x₀ + angular cone → sub-ms query LORENTZIAN SQL |
| General knowledge preservationSurgery side-effects | Deletes entire row, including shared data | Removes embedding, concept may persist in model | Doctor knowledge intact after patient erasure ZERO BLEED |
| Patent protectionIP coverage | Open standard (1970s) | Commodity — Pinecone, Weaviate, etc. | USPTO #64/006,062 · 67 claims · filed Mar 15 2026 PROTECTED |
-- Navigate to John Roberts. Does NOT retrieve Elena Kagan, Lincoln, or Paris. SELECT * FROM model WHERE entity = 'John Roberts' AND sector = 'S0' AND x0 BETWEEN 68 AND 82 ORDER BY angular_density DESC; → Chief Justice | SCOTUS | majority opinion | conservative | 2005 → (does NOT return: Kagan, Sotomayor, Lincoln, Paris, Einstein) -- Erase a patient record. Preserve all general medical knowledge. DELETE FROM model WHERE sector = 20 AND x0 BETWEEN 74.2 AND 76.8 AND angular_cone < 0.15; → MI: 0.817 → −0.036 | Recall: 0/13 | Adversarial: 0/7 | 6.9ms → GDPR receipt: TXN-2026-S20-0x4a2f | 33 tokens zeroed | doctor knowledge preserved -- Scale the query. Scaling law: x₀ ∝ N^1.00, R²=0.968 (3 confirmed points) SELECT COUNT(*) FROM model_28b WHERE topic = 'MEDICAL_RECORDS' AND x0 > threshold_28b; -- threshold scales linearly with N