The Problem
Most AI research tools stop at information aggregation — they collect facts from sources, deduplicate, and summarize. When the novelty rate drops below some threshold, they declare convergence. But knowing all the facts about a domain is not the same as understanding it.
There’s a consensus ceiling: the boundary of what published sources collectively say. Traditional deep research systems can never exceed this ceiling because they only extract and classify existing claims.
G-Infinity was built to break through that ceiling.
How It Works
The system is modeled on educational psychology frameworks — specifically Piaget’s constructivism, Vygotsky’s Zone of Proximal Development, Bloom’s taxonomy, and Meyer & Land’s threshold concepts.
The Core Loop
Each research cycle follows a cognitive construction process:
- Gather — Pull claims from diverse sources across multiple domains
- Construct — Build a mental model by connecting claims into coherent structures
- Challenge — Deliberately stress-test the model by seeking contradictory evidence
- Synthesize — Generate cross-domain insights that no single source contains
- Evaluate — Score the depth of understanding using Bloom’s taxonomy levels
The system doesn’t just collect information — it builds, breaks, and rebuilds its understanding until it can generate genuinely novel insights.
Breaking the Consensus Ceiling
The key innovation is cross-domain synthesis. When Infinity researches neuroscience, it doesn’t just read neuroscience papers. It pulls from cognitive science, philosophy of mind, clinical psychology, and computational theory — then synthesizes connections that individual domain experts typically don’t make.
This is formalized through:
- Gemini embeddings for semantic similarity across claims
- Stats-backed threshold detection to identify when the model has genuinely shifted (not just accumulated more of the same)
- Restructured cross-domain synthesis that forces the system to find structural analogies between unrelated fields
Architecture
The engine is TypeScript-native, designed to run as a long-lived autonomous process:
- Orchestrator — manages research cycles, tracks convergence metrics, decides when to advance stages
- Claim Store — deduplicated knowledge base with embedding-based similarity
- Synthesis Engine — generates cross-domain insights scored by novelty and depth
- Cost Controller — tracks LLM spend per cycle with configurable budgets
Production Numbers
From the neuroscience validation run:
- 3 research cycles
- 64 claims extracted and cross-referenced
- 212 LLM calls
- $0.44 total cost
- Fidelity score: 8.3/10 against the design spec
Lessons Learned
Convergence metrics matter more than you think. V1 used a simple novelty-rate threshold. V2 uses a multi-dimensional convergence measure that tracks not just new information, but new understanding — measured through the system’s ability to generate valid novel claims.
Cross-domain synthesis is where the value is. Single-domain research produces summaries. Cross-domain research produces insights. The architectural decision to force synthesis across domains was the single biggest driver of output quality.
Educational psychology frameworks transfer to AI systems. Piaget’s assimilation/accommodation model maps directly to how a knowledge graph should handle contradictory evidence. Vygotsky’s ZPD maps to how research prompts should be scaffolded. These aren’t metaphors — they’re engineering decisions.
Status
Active development. The v2 engine is validated and producing research outputs. Current focus is on expanding the domain coverage and improving the synthesis pipeline’s ability to identify structurally similar patterns across distant fields.