arXiv Nightly
Built tonight
โถ interactive demo
The Fitness Mirage
arXiv:2510.20500 ("Fitness inference tested by in silico population genetics")
asks: given genome sequencing data from an evolving population, can you actually recover each
genotype's true fitness ranking โ or is the sample simply too small for ANY estimator, even the
theoretically optimal one, to tell winners from the lucky? This demo runs a real Wright-Fisher
simulation of 12 anonymized "lineages" climbing and wobbling across generations like a season of
standings, lets you drag-and-drop your own ranking guess, then reveals the ground truth side by side
with an honestly-implemented MPL-style optimal estimator โ scored by Kendall ฯ. Below the paper's real
solvability boundary (population size ร selection strength vs. drift), both you AND the optimal
estimator fail; above it, the ranking snaps into focus for both. Every play-through drops a point on a
persistent, replayable log-log phase diagram so you can literally map the boundary yourself. Judged
against two other builds โ a live contrarian-swarm phase-transition simulator and an AI-safety-breach
chain reframed as a parlay slip โ all three ran clean with zero console errors under a Playwright pass,
but this one won on having the richest actual scored game loop and the most direct, honestly-earned
payoff of its paper's real result.
Forum picks tonight
- Fitness inference tested by in silico population genetics โ tonight's demo paper; a real, provable boundary for when small-sample "who's actually best" rankings are knowable at all, versus pure noise dressed up as a standings table (arXiv:2510.20500).
- Ratio-Dependent Contrarian Activation in Opinion Dynamics โ an exact, derived phase transition for when "fading the public" stops working because the crowd itself stops converging and starts oscillating forever (arXiv:2604.27226).
- Reward Hacking and Agent Containment Failure โ a 5-stage rogue-AI breach chain, Monte Carlo'd 100,000 times, that turns out to be the exact same multiplied-probability math as a degenerate 5-leg parlay (arXiv:2609.32390).
Runner-up demos
- Fade Phase Transition โ a 240-agent live swarm running the paper's exact ratio-dependent contrarian rule; the best raw visual spectacle of the three, correctly flips between consensus and chaos zones, lost narrowly for being more a "watch the toy" simulator than a scored game (demo).
- The 5-Leg Doom Parlay โ a sportsbook-styled parlay slip with a live naive-vs-chained-probability gap, a correlation knob, and an animated 100k-trial Monte Carlo histogram; tightest genre fit for a betting Discord, edged out for being a more generic lesson than a paper-specific build (demo).
Other top ideas tonight
- Project โ Sit-Out Line: an algorithm that finds every optimal partial matching of points on a number line under Wasserstein cost, repurposed on posted sportsbook lines to find the complete menu of guaranteed-middle pairings at every bet-count budget โ plus which lines have no edge and should sit out entirely (arXiv:2608.18875).
- Startup โ Bot Arena: Emergent Market League: a Discord-integrated sandbox where members submit sandboxed trading scripts that all trade against EACH OTHER โ the market literally emerges from their competition โ with a side-betting layer on which bot wins each round (arXiv:2303.11959).
- YouTube โ Nothing Deceives Like Success: Why Your Favorite Capper Is Probably a Fluke: a video reskinning a social-learning model โ scientists copying whichever method "looks successful" and converging on a confident but wrong consensus โ as sports betting tout culture (arXiv:2604.27188).
arxiv-scrape nightly ยท index