# Falsify

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Do hyun Park](https://cerebralvalley.ai/u/ll11ll)
- **GitHub:** https://github.com/kbns2000/falsify
- **Demo video:** https://youtu.be/6rv56FditVY
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/180

**Falsify — the devil's advocate for market beliefs.**

Every retail trader carries beliefs: "when bitcoin's RSI drops under 30 it always bounces", "three red days means a green one", "a golden cross on DOGE means a month of gains". They are almost never written down precisely, never tested against a fair baseline, and quietly re-tuned after the fact until they look right. Backtesting tools are built to make a strategy look good. Nothing in the stack is built to make a belief *fail*.

Falsify is that missing adversary. You type a belief in any language. Meta's open-weight **Muse Glimmer 30B** rewrites it as one exact rule in a tiny, safe rule language, names the asset, direction, holding period and — crucially — the **pre-registered success criterion**, and lists every assumption it had to make. The rule is then **hashed and chained to a public ledger before a single number is computed**. A deterministic engine replays it on nine years of public candles (signal at the close, entry at the next open, fees on both sides) and runs a fixed battery of attacks: 2,000 random-entry portfolios with the same holding period, fee sensitivity, first half vs second half, nearby parameters (the curve-fitting test), the same rule on other assets, a Bonferroni correction for how many beliefs *you* have already sealed, sample size and a look-ahead guard. Muse Glimmer then plays devil's advocate and proposes counter-tests specific to that belief — the mirror-image trade, a different horizon, the asset you actually trade — which the engine runs under the same rules.

The verdict — FALSIFIED, FRAGILE, SURVIVED (for now) or INCONCLUSIVE — is decided by the numbers, never by the model. The model writes the report in the user's language: what was tested, the strongest evidence against, what held up, what would change the verdict. Every sealed claim and its verdict stay in an append-only, hash-chained ledger that anyone can recompute.

Three design rules make it honest: seal before you see; the model never touches the numbers; a pass is not a forecast. Falsify never tells you what to buy. It tells you whether your belief survived an honest attempt to break it — which is the one thing a trader cannot do for themselves.

Built solo in one week: FastAPI + a deterministic pandas engine, a whitelisted expression compiler, SSE streaming of every stage, hand-drawn SVG charts, 15 tests, one-click Render deploy. Bring your own CSV to test beliefs about any market.

Live demo: https://falsify-baxc.onrender.com — the eight seeded claims and the demo video were produced with Muse Glimmer 30B running locally via Ollama. Without paid inference credits, new beliefs on the hosted demo are formalized by a clearly labelled rule-based fallback; the engine, attacks and ledger are identical. Run it locally with `ollama pull muse-glimmer:30b` for the full model loop.

## More from Meta Global AI Developer Hackathon

- [Godwin Enendu](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/177)
- [Quantum Concepts LLC](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/178)
- [Muse Shopping Index](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/179)
- [AI_NOVA](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/181)
- [Team Been,Meta](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/182)
- [CategoryOs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/183)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/180. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt