fullSET

Tribal knowledge dies here. Heuristics that scale, cite evidence, and survive audit — not slide-deck intuition.

fullSET is a heuristic pattern-synthesis engine implemented in Rust. It consumes typed ProcessedData and emits Insight and Hypothesis values whose confidence is derived from cited evidence and a bounded reasoning chain — not from a model hallucinating plausibility.

Ten engine modules — heuristics, patterns, connections, insights, ontology, semantics, mathematics, integrations, performance, and applications — partition 74,581 lines of production code across 77 source files. Parallel heuristic dispatch with Thompson-sampling adaptive selection. PrefixSpan sequential mining, DBSCAN spatial clustering, and fusion pattern synthesis. An eight-step connection builder constructing intelligence graphs up to 100,000 nodes with five causal-discovery methods: do-calculus, Granger causality, instrumental variables, potential outcomes, and structural causal models.

Advanced mathematics runs in-process: Vietoris-Rips persistent homology, Bayesian network variable elimination, SVD matrix completion, Louvain community detection, and eigenvector centrality. Multi-category insight generation ranks by impact and relevance, deduplicates by similarity threshold, and propagates confidence via multiplicative chains and Noisy-OR combination — every score backed by a parameter-registry row, not a hardcoded constant someone picked in 2019.

fullSET is not your workflow engine, not your knowledge graph server, not your BI dashboard. It is the analysis substrate products compose — binding to athena-engine for ontology, timeline-x for causal lineage on every emitted insight, and the full control-surface stack: CLI, REST, gRPC, Rust/Python/TypeScript SDKs, SQL dialect, and Tauri GUI. Self-hosted, air-gapped, managed cloud, or edge — standard, ultra, and full feature tiers strip or restore GPU, distributed, and advanced-math capabilities at startup probe.

fullSET synthesizes patterns and ranks hypotheses — it does not store durable truth, orchestrate workflows, or sign audit evidence alone. Without ATHENA, ontology queries have no graph to land in. Without timeline-x, insights lose causal_chain_root_event_id provenance and hybrid-logical-clock timestamps. Without KAOS, probabilistic linkage stays advisory instead of entering deterministic merge gates. Without ARGUS, reasoning chains are engineering artifacts, not court-grade proof. Without polyMAX, set context and exchange packages never cross fabric boundaries with typed semantics. Without MOIRA, high-impact insights never trigger governance reactions. Without WFX-b, multi-stage analysis sagas collapse back into one-off notebook runs. fullSET kills tribal heuristics — the other cores kill everything else that would make those heuristics production-safe.

Parallel Heuristics Engine

HeuristicRegistry dispatches Arc<dyn Heuristic> impls concurrently via Rayon — All, TopN, Threshold, and Adaptive (Thompson sampling) selection strategies with ResourceMonitor rejection above 500MB peak.

Domain-Routed Heuristic Dispatch

HeuristicContext filters by Intelligence, Marketing, Enterprise, Analytics, or General domain — context.domain gates eligible heuristics before parallel execute.

PrefixSpan Sequential Pattern Mining

PatternSynthesizer mines sequential patterns via prefix-projected growth — min_support 0.1, configurable prefixspan_min_length 2 and max_length 20.

DBSCAN Spatial Pattern Detection

Density-based spatial clustering (Ester et al. 1996) identifies spatial anomalies and region patterns from ProcessedData with similarity_threshold 0.7 default.

Eight-Step Connection Builder

ConnectionBuilder constructs ConnectionGraph values — entity-to-entity strength, topology analysis, transitive closure, batch/chunk parallel construction up to 100K nodes and 4096MB memory limit.

Five-Method Causal Discovery

Do-calculus, Granger causality, instrumental variables, potential outcomes, and structural causal models — each parameter-registry governed with alpha, max_lag, bootstrap iterations, and cross-validation fold defaults.

Multi-Category Insight Generator

Pattern, causal, predictive, anomaly, trend, recommendation, and composite insights — MultiFactor ranking via w_impact * impact + w_relevance * relevance with Jaccard deduplication above diversity_threshold 0.3.

Confidence Propagation Engine

Multiplicative chain combination and Noisy-OR for independent evidence — chain_penalty_base 0.95 per step, min_confidence 0.5 gate before surfacing results.

Natural-Language Explanation Generator

Reasoning chains convert to operator-readable explanations with Brief, Standard, Detailed, and Comprehensive detail levels — every insight requires reasoning_chain.len() >= 1.

Topological Data Analysis

Vietoris-Rips complex construction, boundary matrix reduction, persistence diagrams, Betti numbers, and bottleneck distance — parallel construction via Rayon with persistence_threshold 0.1 default.

Bayesian Network Inference

Variable elimination over conditional probability tables with full factor algebra — Koller and Friedman 2009 reference implementation in mathematics module.

SVD Matrix Completion

Sparse matrix reconstruction via singular value decomposition with convergence tolerance from parameter registry — Candes and Recht 2009 completion theory.

Louvain Community Detection

Greedy modularity maximization (Blondel et al. 2008) partitions connection graphs — community insights carry reasoning chains of length >= 2 citing modularity evidence.

Ontology Interface via AthenaConnector

Concept mapping, taxonomy navigation, find_concepts, find_related_concepts, and semantic_search against athena-engine — SPARQL-style querying with per-endpoint cache TTL and exponential backoff retry.

Semantics Interface (NLU Pipeline)

Entity recognition, sentiment analysis, topic modelling, coreference resolution — BERT-base-uncased default with 512-token max sequence, four-language support, 0.7 confidence threshold.

Inference Rule Fixed-Point Engine

Condition evaluation over Comparison, Or, and Not branches with fixed-point iteration — rule engine produces Hypothesis values with cited InferenceStep chains.

Timeline-X Causal Lineage Binding

17 integration call sites — every Insight carries causal_chain_root_event_id, HLC timestamps, and ingestion bounded by 256-permit semaphore.

132-Row Parameter Registry

Engine, heuristics, patterns, connections, causal, insights, ontology, semantics, performance, and timeline-x parameters — SET/SHOW via SQL dialect and CLI config commands.

Full Control-Surface Stack

CLI with psql-style backslash commands, OpenAPI 3.0.3 REST, tonic gRPC/protobuf, Rust/Python/TypeScript SDKs, PostgreSQL-compatible SQL dialect, Tauri GUI, and health endpoints — config-injected base URLs per deployment.

Tiered Feature Architecture

standard, ultra, and full Cargo feature tiers — edge startup probes detect cores, memory, and accelerators then apply tier defaults; air-gapped builds strip outbound networking and telemetry phone-home.

See it ignite.

A demo of one core is a tech talk. A demo of all nine is a reckoning. Book the second one.

Get in touch

fullSET

Tribal knowledge dies here. Heuristics that scale, cite evidence, and survive audit — not slide-deck intuition.