Our Team

Chainscore Labs is a specialist Web3 R&D firm with deep expertise in blockchain infrastructure, distributed systems, and AI. Our team combines seasoned software engineers (Python, Rust, smart contracts) with AI researchers experienced in symbolic reasoning, probabilistic logic networks, and meta-programming. 

Prasad - Lead: 6+ YOE in Web3 R&D and protocol design
Siva - Academic background in cryptography and blockchain Committed to rigorous empirical validation

Company Name (if applicable)

Project details

Conceptorium extends Hyperon’s concept formation capabilities by integrating two methods: Contextual Conceptual Blending and Uncertain Formal Concept Analysis (FCA) - as native MeTTa modules operating over the Distributed Atomspace. Rather than relying exclusively on simple clustering, Conceptorium generates new ConceptNodes that reflect both creative combinations of existing knowledge and structured generalizations tolerant of uncertainty and contradiction.

Importance and Gaps Addressed: Hyperon’s current concept induction produces static groupings that lack sensitivity to the agent’s active goals and environmental cues. Additionally, it does not gracefully handle noisy or conflicting data, nor does it offer automated mechanisms to assess the quality of newly formed concepts. Conceptorium addresses these shortcomings by enabling context-aware creativity, robust handling of imprecise information, and a built-in evaluation pipeline that filters for coherence, novelty, and practical utility.

Core Mechanisms

Contextual Conceptual Blending
This module accepts two or more source ConceptNodes and operates within a specified context subgraph defined by attention weights. It first aligns the structural roles and relational patterns of the inputs, then selects salient feature subsets from each. Conflicting attributes are resolved either by computing a weighted average of fuzzy truth values or by marking them with paraconsistent truth-pairs when direct contradiction occurs. Each candidate blend is scored according to conditional mutual information—which measures the unexpected overlap of features within the context—and the amount by which adding the blend would reduce overall Atomspace complexity, minus any conflict penalties. The top-scoring ConceptNodes are promoted for integration.

Uncertain Formal Concept Analysis (FCA)
The FCA module transforms object-attribute associations within a context subgraph into either real-valued membership degrees (fuzzy) or dual truth-pairs (paraconsistent). It then applies a customized concept lattice algorithm—an adaptation of the NextClosure procedure—to enumerate formal concepts whose extents (sets of objects) and intents (sets of attributes) meet predefined support and size thresholds. Each extracted concept is represented as a ConceptNode with graded membership links and is systematically inserted into the hierarchy via inheritance relations, forming a partial order that captures nuanced generalizations even in the face of uncertain or contradictory evidence.

Processing Workflow Upon activation, Conceptorium’s controller gathers the relevant context by querying Atomspace for ConceptNodes currently weighted above a threshold by the Economic Attention Network (ECAN) and their immediate neighbors. The Contextual Blending and FCA modules then execute in parallel on this context subgraph. Once candidate ConceptNodes are generated, the Evaluation pipeline assesses each using symbolic metrics—such as entropy reduction and feature coherence—supplemented by PLN-based surprise detection that runs forward-chaining to surface unexpected inferences. An optional language-model-based component can further rate novelty by soliciting structured feedback from multiple LLM agents. Finally, only those concepts meeting overall score criteria are atomically committed back to Atomspace, and ECAN is notified to adjust importance values accordingly.

MeTTa API Conceptorium exposes a concise set of MeTTa primitives for ease of use:

• concept_blend(sources, context, params): Executes Contextual Blending on specified source ConceptNodes within the given context, returning new blends.

• run_fca(objects, attributes, params): Performs Uncertain FCA on the provided object-attribute matrix, returning newly discovered concepts.

• evaluate_concept(node, context, params): Computes a combined quality score for a single ConceptNode under a specific context.

• list_contexts(): Retrieves all active context frames.

• create_context(params): Defines a new context frame for targeted concept formation.

Evaluation Metrics Conceptorium validates its output using four key metrics: the Novel Concept Rate (the percentage of candidates that pass filtering), computational performance (average time per concept and throughput), quality (the composite evaluation score combining symbolic, PLN, and optional LLM-based assessments), and reasoning impact (measured improvement in PLN inference efficiency when using the new concepts).

Scalability and Reliability To support large-scale knowledge graphs, Conceptorium parallelizes concept generation tasks across Atomspace shards and reuses cached computations for unchanged context regions. Under heavy load, it dynamically tightens thresholds for feature selection and concept support. Each module is wrapped in error-handling logic to log failures and continue processing unaffected contexts, ensuring the system remains responsive even when individual operations encounter exceptions.

Illustrative Examples When operating in a "study" context, blending the concepts Library and Cafe produces a ConceptNode "Library-Cafe," which combines quiet reading spaces with beverage service. In a gaming scenario, applying uncertain FCA to character trait data yields archetypes such as "Guard-Diplomat," capturing entities that are both cautious and socially adept.

Future Enhancements Planned extensions include reinforcement-learning-driven selection policies that learn which blends succeed most often, integration of a simulation engine to validate the physical plausibility of certain blends, and automatic predicate invention via information-theoretic detection of frequently co-occurring feature sets.

Conceptorium transforms Hyperon’s concept formation from a basic clustering mechanism into a full-fledged, context-aware synthesis engine. By combining creative blending with rigorous formal analysis under uncertainty, this project equips AGI researchers with powerful tools to explore, evaluate, and integrate novel abstractions directly within the MeTTa/Hyperon ecosystem.

Links and references

Pitch Deck - https://drive.google.com/file/d/1YrLo2VKsk9Ga8Iy2s0Zv1srWEjw5afdq/view?usp=sharing