Aurora

Aurora is the verification layer for machine-produced quantitative claims I’ve been calling Aurora a “glass-box quantitative intelligence engine” since launch. That was never wrong, but it described the machinery instead of the job. Here’s the sharper version: Aurora is the verification layer for machine-produced quantitative claims. Not an analyst — the thing that checks the analyst. Here’s the problem it exists for. An AI gives you a number. Maybe it computed that number properly. Maybe it produced something plausible-shaped, because plausible-shaped is what the training rewards. You can’t tell by looking, and the more fluent models get, the less you can tell. Aurora takes the claim and the data and comes back with one of three answers: supported, contradicted, or not identifiable. Every finding traces to a named method. Every run produces an integrity-hashed bundle that someone else can verify without having to trust you or me. That third answer is the one I care most about. When the data genuinely can’t settle a question, Aurora says so instead of returning a confident number. It was the hardest thing to build, and it’s the reason the other two answers mean anything. Nothing about the product changed with this. 19 research-grade methods, causal inference with do-calculus, live streaming mode, Decision Contracts, signed bundles — all shipped, all working. What changed is that I stopped describing what’s under the hood and started describing what it’s for. pip install aurora-mcp Live on PyPI, Glama, and the official MCP registry. Point Claude Desktop, Cursor, or any MCP client at it and your agent gets methods it can call instead of numbers it has to guess at. Still local-first. Still no telemetry. Still Apache 2.0. What’s next: calibration. Aurora should know how often each method is wrong on data shaped like yours, and report that alongside the finding.

Your agent is guessing at numbers. Aurora is one pip install away from stopping that. pip install aurora-mcp Live on PyPI, Glama, and the official MCP registry as of this week. Point Claude Desktop, Cursor, or any MCP client at it and your agent gets 19 research-grade statistical methods it can call instead of estimating through — anomaly detection, changepoints, causal inference, spectral analysis, the real ones. Every finding comes back cited, traced to the method that produced it, wrapped in an integrity-hashed bundle you can hand to someone else and they can verify. Fabricated count: 0, enforced by a validation gate rather than a prompt. And when the data genuinely can't answer the question, it says "not identifiable" instead of returning a confident number. That turned out to be the hardest feature to build and the one I'm proudest of. Also shipped since the last update: live streaming mode with contract alerts to Slack/Discord/email, a plugin SDK for your own methods, custom PDF ingestion into the knowledge bank, a runs library with A/B comparison, and Docker self-hosting. Test suite went from 320 to 699. Still local-first, still no telemetry, still Apache 2.0. pip install aurora-mcp github.com/FantasyLab-ai/aurora What's the number your agent got wrong that made you not trust it again?

Aurora since launch day: live streaming mode, real causal inference, a plugin SDK, and 379 new tests Last update here was the Product Hunt launch. Aurora hasn't stopped since — the whole v1.2 sprint is on main and most of v2.0 landed ahead of schedule. The big ones: 🔴 Streaming mode — point Aurora at a file or folder and it watches, runs incrementally on a rolling window, and fires Decision Contracts on genuinely new findings only (deduped, so it's alerts and not spam). Live findings strip in the Studio, SSE event feed if you want to wire it yourself. ◎ Real causal inference — do-calculus with backdoor identification, plus Pearl-style counterfactuals. On a synthetic confounded dataset it recovers the true effect within ~0.05. When an effect isn't identifiable from the data, it tells you that instead of returning a confident number. 📊 24+ methods, up from 17 — VAR, dynamic time warping, Bayesian online changepoint detection, robust PCA, empirical mode decomposition, Kalman, spectral entropy. Each gets its own tile, each with honest skip reasons and crash isolation. 🔌 Plugin SDK — write your own methods and Aurora runs them alongside the built-ins. The validation gate hard-refuses any finding flagged fabricated. You can extend it without being able to break the honesty contract. 📚 Bring your own knowledge — ingest your own PDFs and folders into the knowledge bank, plus starter packs for Climate / Finance / Biomed / Industrial. 🔔 Contracts that actually reach you — Slack, Discord, and email actions alongside the generic webhook. TLS-only SMTP, credentials from env vars, never from the contract file. 🗂 Runs library — pin runs, A/B compare any two (new vs. disappeared findings), export a portable signed bundle. 🐳 Self-hostable — Docker + compose, pluggable LLM providers (Anthropic / OpenAI / Gemini / Ollama / local), 🧪 699 tests passing locally, up from 32

Officially launched on Product Hunt! Check it out now! https://www.producthunt.com/products/aurora-22?utm_source=other&utm_medium=social

Progress update on Aurora, the local-first, glass-box quant tool (drop a CSV → cited findings from real statistical methods, 0 fabricated). Since the last drop: 🔎 A real data workbench — every run now has an interactive correlation matrix (Pearson/Spearman heatmap), click-to-scatter with regression + r/r²/p/CI, plus Distributions and Quality tabs. Explore your data without leaving the app. 🎯 A decision layer — this is the part I care about most. Findings were the "what"; now every finding carries a recommended decision, the risk if you ignore it, and a one-click decision brief you can paste into Slack or a doc. From insight → action. ◎ "Is it real?" — on any correlation, run do-calculus to check whether it survives as a causal effect. Correlation is the hook; causation is the payoff. (Also fixed a bug that had the whole causal engine silently returning errors — oops, but caught it.) 🖥️ Still the same core promise: 100% on your machine, no API keys, no cloud, no telemetry. Runs a real battery of methods — changepoints, HMM regimes, anomalies, matrix profile, VAR, Kalman, robust PCA, ensemble forecasting, and SINDy physics discovery (it literally recovers governing equations from data). Try it: free download (Windows / macOS / Linux) → github.com/FantasyLab-ai/aurora/releases A free, no-install web demo is landing next so you can see it work in one click. Building in public — feedback (and upvotes 🙏) very welcome. What would make you actually trust an automated finding?

Quick update on Aurora, the glass-box quant tool. Just cut v0.2.0 of the desktop app — it's now a true download-and-run experience: the analysis backend is bundled inside the installer, so no Python, no terminal, no API keys. New in the desktop: Causal Lab — interactive do-calculus. Pick a treatment + outcome and get an effect estimate, standard error, and the adjustment set it actually used. When the data can't identify the effect, it tells you it's unidentifiable instead of guessing. Phase Space — the system's trajectory + attractors, rebuilt native (was a legacy graph). Spacetime — per-entity timelines with a scrubber. Forecast — forecasts with confidence bands and user inputs. Knowledge Bank — browse/search the cited evidence base (per-source breakdown + free-text search). Plus: first-run onboarding, a persistent run-context bar, and insight-style finding cards (severity + evidence chips + "cited" badges). Fixed a "results show one run behind" race, and got cross-platform release CI green: Windows .exe/.msi, macOS .dmg, Linux .deb/.rpm. Same core promise as always: every finding is tied to the actual statistical method, the fabricated-claim count is surfaced (and stays at 0), and it says "I can't answer that" rather than inventing a number. Download: https://github.com/FantasyLab-ai/aurora/releases/latest Feedback very welcome — especially on the Causal Lab UX.

Big milestone — and it's the one that was quietly blocking everything else. Until now, running Aurora meant cloning a repo, setting up Python, installing dependencies, and launching a server from the terminal. Fine for developers. A brick wall for everyone else. That wall is gone. Aurora is now a native desktop app you install like any other software. Download it, double-click, and it opens — analysis engine and all. No Python. No terminal. No setup. What shipped: 🖥️ A real desktop app — frameless native window, sidebar navigation, drag-and-drop (drop any CSV, JSON, Parquet, or Excel file right onto it). Findings you can click into for the evidence, a plain-English narrative of what Aurora found, methods breakdowns, and your full run history. 📦 The whole engine, bundled inside. This was the hard part — freezing Aurora's entire scientific stack into one installable app took real surgery. But it works: I verified a clean install runs a full analysis offline, writing cited results with zero fabricated numbers. 🌍 One-click installers for Windows, macOS, and Linux. The point of Aurora has never changed: a local AI that computes instead of guesses, and cites every number it gives you. What changed this week is that you no longer need to be a developer to use it. Installers are landing on the releases page now. If you try it, I genuinely want to hear what breaks and what you want next. Built in the open, by one person, for real work — your feedback steers it. More soon. 🐻 — Brandon / FantasyLab