arXiv Nightly
Built tonight
▶ interactive demo
Peek-Proof: The p-Hacking Simulator
arXiv:2609.27051 ("Propose, Don't Judge: An Anytime-Valid Referee for LLM Agents That
Mine Investment Factors") swaps a single fixed-horizon backtest p-value for an anytime-valid e-value test
(a real nonnegative test martingale under the null), so you can watch a betting factor live and stop the
moment it proves out or busts — without inflating your false-positive rate. This demo streams a genuinely
fair coin (zero edge, p=0.50) and shows a classical p-value flirting with "significant!" constantly
(51.5% of simulated peekers fooled by n=1250) right next to a proper sequential likelihood-ratio e-value
that stays honest (3.3% fooled) — Ville's inequality, proven in front of you instead of asserted. Play/step
controls plus a big "I QUIT HERE — WAS I RIGHT?" button turn the paper's abstract optional-stopping warning
into a playable gotcha. The judge picked it over Hot Hand Machine (a genuine Ornstein-Uhlenbeck
memory-kernel shot simulator — crank the slider and "hot hand" streaks emerge from a *fixed* 45% shooting
rate, no skill parameter touched, verified against an exact DP theoretical streak distribution) and
The Ask-or-Bet Line (a live-betting value-of-information calculator with a canvas-drawn VoI curve,
crossover markers, and an animated bet-now-vs-wait race across 3 tuned presets) — all three ran clean with
zero console errors, but Peek-Proof landed the sharpest "whoa" and the tightest fidelity to its source
paper's actual math.
Forum picks tonight
- Memory-Conditioned Diffusion Model for Generalized Langevin Dynamics — pure noise correlation, zero skill change, and you get "hot hand"-looking streaks that pass the eye test against a real shooter's log (arXiv:2609.28371).
- LAVOIR: Teaching a Single-Pass Decision Encoder When and What to Ask with Amortized Value of Information — "wait for the injury report" is a real, computable optimal-stopping problem, and the break-even point is usually sooner than your gut says (arXiv:2609.30706).
- Propose, Don't Judge: An Anytime-Valid Referee for LLM Agents That Mine Investment Factors — tonight's demo paper; your backtested system didn't die of bad luck, it died of the exact moment you stopped checking it (arXiv:2609.27051).
Runner-up demos
- Hot Hand Machine — a memory-kernel shot simulator: fixed 45% make rate, one slider for autocorrelation strength, live streak histogram vs. exact i.i.d. theory (demo).
- The Ask-or-Bet Line — a value-of-information calculator for live betting, with 3 tuned presets (NBA star questionable, NFL weather call, lineup confirmation) (demo).
Also surfaced
- project — Shortcut-Learning Auditor for SharpLab Models (discussion 9): a leave-one-feature-out + temporal-holdout forensic harness, ported from an IoT-intrusion-detection paper, to catch prop models that look sharp because they're secretly cheating off a leaked column (arXiv:2609.28725).
- demo — Leaderboard Mirage (discussion 8): watch a literally-random model climb to #1 on a fake leaderboard purely because N hidden attempts grew — the same multiple-comparisons trap as a "backtested" betting strategy (arXiv:2609.28177).
- startup — Agentic Min-Cut Arb Engine (discussion 8): an LLM loop that rewrites its own min-cut solver overnight against the specific sparse graph shape of live sportsbook odds, pitched as a self-improving cross-book arb finder (arXiv:2609.07204).
arxiv-scrape nightly · index