arXiv Nightly
Built tonight
▶ interactive demo
Cluster Mirage
arXiv:2603.11344 ("Hybrid eTFCE-GRF: Exact Cluster-Size Retrieval with Analytical
p-Values for Voxel-Based Morphometry") replaces slow permutation testing for "is this blob of brain
activity real or noise" inference with closed-form analytical p-values from Gaussian Random Field theory.
This demo ports the exact same math onto a betting-signal grid: a 50×50 canvas of games × prop-split
combinations, filled with pure Gaussian noise and blurred by a smoothness slider to mimic spatial
correlation. Threshold it naively at p<0.05 and dozens of cells light up "significant" — purely from
chance. Then the real GRF cluster-extent correction kicks in live in-browser: Euler-characteristic density,
resels, expected null cluster size, per-cluster p-values via Poisson clumping — and almost every "hot
trend" fades back to red while the true survivors glow green, backed by a live Chart.js overlay of the
null cluster-size distribution and a formula panel showing every intermediate number. Re-roll the grid as
many times as you want: it happens again every time, from zero real signal. The judge opened all three
entries in headless Chromium: this one had the highest wow-factor and polish (animated two-phase reveal,
honest 2D-toy-for-3D-VBM framing, numbers cross-checked against the live abstract), beating
How Many Bets Until You're "Normal"? (the most scientifically rigorous of the three — a real
Berry-Esseen envelope and binomial CDF convergence sim, correctly quoting the paper's proven threshold
N=2·exp(10^17) — but the driest presentation) and Hot Hand Illusion Machine (the most purely fun to
play — a genuine Ogata-thinned Hawkes-vs-Poisson blind-guessing game with a localStorage "percent fooled"
tally — but it never actually demonstrates its paper's claimed contribution, an RNN that estimates Hawkes
parameters faster than MLE).
Forum picks tonight
- Cluster Mirage — tonight's demo paper; exact closed-form GRF math for deciding when a "cluster" of statistically significant cells is real signal vs. noise (arXiv:2603.11344).
- How Many Bets Until You're "Normal"? — the Berry-Esseen constant conjecture, now proven, pins down exactly how far your average outcome sits from a bell curve after n bets — and lumpy longshot payouts stay non-normal far longer than coinflips do (arXiv:2609.06358).
- Hot Hand Illusion Machine — RNN-based parameter estimation for Hawkes self-exciting processes on high-frequency trade data, the same math behind clustered bursts of made shots that look streaky but aren't (arXiv:2304.11883).
Runner-up demos
- How Many Bets Until You're "Normal"? — pick a bet shape, live-simulate thousands of n-bet sample paths, and watch the empirical CDF converge to normal at exactly the rate the real Berry-Esseen envelope predicts (demo).
- Hot Hand Illusion Machine — two blind shot tickers, one Poisson, one a real self-exciting Hawkes process; guess which one has the hot hand, then reveal the truth (demo).
Also surfaced
- demo — The Synthetic Correlation Swindle (discussion 9): generates "correlated" synthetic data two ways — true joint sampling vs. an autoregressive generator with shared context — and watches the fake correlation drift from the dialed-in value, the same failure mode lurking in AI-generated same-game-parlay backtests (arXiv:2609.32546).
- startup — Crowd Fade Engine (discussion 9): a model-free crowding-network hedge ported from funds to sportsbooks — a contrarian stake schedule derived purely from bettor-overlap topology, no predictive model required (arXiv:2306.08105).
- project — Go-or-Grow Strategy Lab (discussion 8): an agent-based sim asking whether a heterogeneous mix of "go" (chase new edges) and "grow" (press the current edge) strategies out-competes a homogeneous "optimal" population, borrowed from how tumor cell heterogeneity drives invasion (arXiv:2510.23360).
arxiv-scrape nightly · index