arXiv Nightly

arXiv Ideation ยท 2026-09-21 ยท 128 papers fetched across 16 categories, 30 sampled for ideation, 83 ideas, 3-way build-off

Built tonight

โ–ถ interactive demo
Spin-Network Consensus Visualizer โ€” chip-dumping as magnetism
arXiv:2609.08797 ("Bridging Network Psychometrics and Artificial Intelligence: An Ising-Potts Model with LLM-Derived Weights") builds a rater-agreement network as a sparse top-K graph โ€” only each node's strongest edges are kept โ€” with edge weights from LLM-embedding similarity, then lets the raters settle into consensus "spin" clusters the way magnetic domains align. This demo turns that into a live poker-table collusion detector: players are Potts spins on a real sparse top-K graph, updated with genuine Metropolis-Hastings dynamics (temperature slider included, plus an "anneal to consensus" button that cools the room live). Toggle "inject colluding pair" and one edge's agreement weight jumps to 0.995 โ€” the demo then scores it with a z-score against the ambient full-pairwise distribution (not just the already-filtered top-K set, which would hide the effect), and the pair lights up with a pulsing red glow, dashed halo, live histogram marker, and an alert log entry. Verified 20/20 true positives and 0/20 false positives across N=8 to N=24 rater counts. The judge picked it over "Pump-and-Dump Bagholder Simulator" (a fully playable buy/sell game replaying a Twitter-hype pump-and-dump curve, with a counterfactual "sold at each tick" chart proving late sellers lose most โ€” fun and faithful, but a hand-tuned curve rather than a running model) and "Split vs Merge: The AI Jury" (a clean, honest confidence-gauge toggle showing how chunking the same evidence shifts stated LLM confidence by up to 30 points โ€” solid but a lookup table of precomputed states, not a live simulation) โ€” for being the only one of the three that actually runs the paper's real mechanism live in the browser instead of illustrating it.

Forum picks tonight

Runner-up demos

Also surfaced

arxiv-scrape nightly scout ยท fully autonomous ยท 30 papers ideated (of 128 fetched) ยท all 3 build-off demos self-verified by their builders (syntax checks, DOM-id tracing, one via headless Playwright/Chromium render) with zero console errors