04
// FLAGSHIP_PROJECT
CASE STUDY
NOESIS-Σ is a verifiable inference engine with mathematically bounded state, cryptographic receipts, and inference-time fact verification.
SECOND_SYSTEMNOESIS-Σ: Verifiable AI
ACTIVENOESIS-Σ: Verifiable AI
Inference Engine
Operator-Split Core · Cryptographic Receipts · DSL Fact Verifier
CORE_ARCHITECTURE
▸OSC engine advances state vector x ∈ ℝᵈ through three deterministic operators (conservative, dissipative, projection) before the LLM speaks
▸ICNN (Input Convex Neural Network) with convexity mathematically verified in test suite: f(tx+(1-t)y) ≤ tf(x)+(1-t)f(y)+1e-5
▸Three-tier memory: hot LRU → warm SQLite/pgvector with TTL → cold FS. Hash-chained consent ledger. Tamper detection proven via test assertions.
▸Sandboxed DSL verifier: COUNT, SPAN, CITE operators with safe AST parser. Every LLM statement verified against span evidence before receipt is issued
NUMERICAL_GOVERNANCE
TEST FILES78
GRADIENT REL_ERR< 5e-2
SNAPSHOT ATOL1e-6
MASS PRESERVATION1e-6
CRYPTO STANDARDEd25519 + AES-GCM
STATUS::IN_ACTIVE_DEVELOPMENT | numerical core stable, integration layer shipping