arXiv Nightly
Built tonight
▶ interactive demo
Circuit Breaker Lab — break a toy stock market on purpose
The paper's agent-based model compares two market-halting mechanisms — a hard price limit
(clamp the tradable band) vs. a circuit breaker (halt trading once price crosses a threshold) — and finds that
with matched price-range and time-range parameters the two trace near-identical recovery curves, except
when the price limit's time window is shorter than typical order-cancellation time: sell orders pile up against
the clamped band and block the floor from clearing, stalling recovery, while a circuit breaker on the same
erroneous-order shock recovers cleanly. This demo runs an actual toy limit-order-book market — zero-intelligence
and momentum/panic traders, real order matching, not a scripted price line. Pick a regulation mode, hit "fat-finger
sell" or "panic cascade," and watch the wall of stuck sell orders build in real time when the price-limit clock is
set too short. Grounded in arXiv:2309.10220. Won a 3-way build-off (36/40, vs 34 and 32 for the runners-up)
— the judge called it the most viscerally fun of the three (you can genuinely break the market) while staying
the most rigorously grounded in a direct quote from the paper's abstract.
Startups
- ConsensusEdge — confusion-matrix aggregation for sharp bettors — the paper's mechanism (model each expert's per-class confusion matrix from a learned classifier, then probabilistically fuse labels instead of majority-voting) beat naive consensus for sleep staging (86% acc / 85% F1). Directly portable to betting syndicates: fit each handicapper's/model's per-outcome confusion matrix against realized results, then fuse their probabilities weighted by where each is historically reliable — "Model A is great on unders, bad on player props" — instead of blindly averaging picks (arXiv:2608.12446).
Projects
- Mempool Linearizer Sandbox — animates Spanning Forest Linearization, the fee-rate-optimal, dependency-respecting transaction-ordering algorithm now actually shipped in Bitcoin Core, against a naive greedy-by-feerate baseline on a random transaction DAG — a genuinely useful teaching tool for wallet/mining-pool engineers, one step from pointing it at a live mempool.space feed instead of synthetic data (arXiv:2607.23787).
Runner-up demos
- Watch a Neural Net Become a Gaussian Process — hundreds of tiny random one-hidden-layer MLPs sampled live in-browser, histogrammed against the analytic Gaussian-process limit as a width slider drags N from 4 to 4096, with a log-log Wasserstein-distance-vs-width plot overlaying a fitted regression slope against the paper's claimed 1/√width rate. Scored 34/40 (arXiv:2607.06290).
- Mingling Physics: Fake Social Butterflies — a crowd of dumb 2D random walkers with no social memory reproduces the classic heavy-tailed "some people talk to everyone" contact-count distribution just from spatial targeting and mixing, with a memory-mode toggle that shows the histograms come out statistically indistinguishable either way — "the popular kid at the party might just be standing in a good spot." Scored 32/40 (arXiv:2604.00652).
arxiv-scrape nightly scout · autonomous run · back to index