The philosophy is strict. Answers are never hardcoded, and neither are the facts behind them. Knowledge lives as data: WordNet, ConceptNet, Wikidata, Wikipedia, Wiktionary, curated schemas, and library API descriptions become nodes in a graph the engine reasons over. If Aria cannot derive an answer, it says so. Zero hallucination is treated as more important than answering everything, and every conclusion should be traceable from question to meaning to graph lookup to reasoning to answer.
The workspace is a Rust monorepo, around 45,000 lines, split across graph storage, a compiler, a reasoning engine, memory, verification, a runtime, CLI, HTTP server, symbolic math, code analysis, code generation, and knowledge ingestion. The knowledge base sits at roughly 550,000 nodes across the sources above. A recent overhaul replaced fixed responses with a compositional natural-language generator: identical facts, freshly worded each time, so the output varies without the underlying knowledge ever drifting.
Code generation has moved from generating from predefined IR specifications to graph-driven synthesis. Aria treats libraries as knowledge: imports, methods, usage, and relationships are graph data, and programs are composed from that structure rather than remembered from training. The same shift is happening for website generation: knowledge to layout reasoning to section reasoning to content generation, with no fake testimonials, fake users, fake statistics, or fabricated companies.
Development runs as a continuous loop of finding failures, isolating root causes, planning, implementing, verifying, and stress-testing, with a custom benchmark inspired by SWE-bench that scores resolution, correctness, engineering quality, efficiency, generalisation, reliability, and difficulty rather than a pass or fail. Unit tests grew from 466 to 499 in the latest pass with zero regressions.
Where Aria is strong today: symbolic reasoning, graph traversal, deterministic mathematics, explainability, code analysis, factual honesty, structured knowledge retrieval, consistent outputs, and the zero-hallucination policy. Where it is weaker, on the record: novel software architectures, large interactive systems like games and GUIs, high-level planning for very large projects, and semantic intent extraction. These are treated as engineering work, not hidden shortcomings.
The long-term goal is not to compete with LLMs on next-token accuracy. It is an AI that genuinely understands language, reasons symbolically, grows its own knowledge graph, generates original software from understanding rather than imitation, explains every conclusion, never fabricates, and improves deterministically over time. Not shipping publicly yet. When it's good enough, I'll link it here.
