Aurora

Aurora is an open-source quantitative intelligence engine for people and AI systems that can’t afford to hallucinate. Most AI tools generate confident-sounding numbers and conclusions, and you usually can’t tell which are real and which are invented. Aurora is built on a structurally different approach: classical statistical methods do the actual computation, and a local language model only narrates what was computed, grounded in real citations. The tagline says it plainly — cloud LLMs guess, Aurora computes. Drop in a dataset (CSV, Parquet, JSON, or XLSX) and Aurora runs 17+ research-grade methods — Isolation Forest, robust z-score, Granger causality, HMM regime detection, SINDy, persistent homology, Gaussian processes, mutual information, and more. It surfaces anomalies, tests causal relationships, detects regimes, and produces forecasts with confidence bounds. Every finding is a structured object with a method, severity, threshold, and a citation that links to the exact knowledge-bank entry behind it. The synthesis paragraph reads like a research abstract because each sentence traces to a published source — Newton, Granger, Hampel, NOAA, NIST, and others. No invented numbers, no invented papers. A live “0 fabricated” indicator is the contractual signal that every claim is grounded. Aurora has two faces sharing one engine. Copilot is the local studio for analysts, quants, scientists, and engineers — six analytical lenses, a spacetime system graph, and phase-space projection for exploring findings visually. Cortex is the verification layer AI builders call when their agents need cited math instead of guesses, available through a Python SDK, an MCP server (works with Claude Desktop, Claude Code, Cursor, and custom agents), and Decision Contracts that fire webhooks or actions when findings match defined conditions. Every layer produces a portable, signable .aurora.json bundle with SHA-256 integrity and optional Ed25519 signing, so a result can be verified on any machine and audited for tampering. Core principles: glass-box at every layer, local-first always (your data never leaves your machine, no telemetry, no phone-home), an honesty rule that renders uncertain findings as uncertain and discloses any skipped or sampled methods, and full open source under Apache 2.0 — inspectable, forkable, and free to use commercially. A typical run completes in around 14 seconds on consumer hardware, fully local after the initial knowledge-bank download. The project ships with honest documentation of what’s solid and what’s still rough, a public roadmap, and a build-in-public log. Aurora is part of FantasyLab.ai, a set of local-first AI tools for serious work, built by Brandon Grutkowski. It’s free, open source, and runs entirely on your own machine. GitHub: github.com/FantasyLab-ai/aurora