Welcome to Shelfware Labs

Sometime in the summer of 2026 I was looking at a market on Kalshi — an exchange where you bet real money on whether things will happen — titled "Best AI this month?", with four contenders: Claude, Gemini, ChatGPT, and Qwen. On an exchange like this, a bet pays you exactly one dollar if you're right and nothing if you're wrong, and for each model you could bet Yes (this one wins) or No (this one doesn't). Then I noticed something that felt illegal. Exactly one model can win the month, so if you bet No on all four at once, three of those bets are guaranteed to come true — three of the four are certain to lose — paying a dollar each. Three dollars, locked in. The four No bets together cost just 298¢. Two cents of free money, sitting in public, on a regulated exchange.

I did what any reasonable person does with two cents of free money. I opened a chatbot and asked it to help me build a machine to find more of it.

What the chatbot did next is the reason this website exists. Instead of cheering me on, it wrote a meticulous, seventeen-hundred-line plan — and folded into the plan were all the ways it would fail. Treat tiny post-fee edges skeptically. "Mutually exclusive" is not the same as "exhaustive." A high annualized return is not the same as real money. Don't build the bot until the scanner proves there's anything to catch. In the sports appendix it went further and simply stated the ending: "Kalshi is probably not globally soft."

I ignored all of it and set out to confirm every word — one edge at a time, across every market I could point a scanner at.

The scanner got built — real .NET, tested to the penny, pointed at 28,900 markets. It found 293 "arbitrages," and every single one was a fee illusion, a mispriced tail, or one contract deep. So I went looking for a prediction edge instead. I checked Kalshi against Pinnacle, the sharpest sportsbook alive, across four sports — Kalshi was just as sharp. I tried to out-forecast its weather markets with five models; the market forecast better than my meteorology. I reconstructed the entire Polymarket leaderboard off the blockchain to copy the winners, and discovered there weren't any — just a swarm of market-making bots harvesting rewards at a one-cent spread, a great many people who had flipped ten heads in a row, and a few genuine sharps whose edges were already expiring — who, in a twist I'll come back to, all turned out to share a single bankroll. I tried to market-make the weather and hit a wall made of arithmetic. I hunted cross-exchange arbitrage and found twenty-one of them, all lies.

I call it the Referenceability Law, and the rest of this site is the evidence.

Shelfware Labs is where those investigations get written down — honestly, with the numbers that killed them, and with enough distance to find them funny. Shelfware is software that gets built and never used. Everything here is shelfware. I'm not embarrassed about it; I'm documenting it, because a well-run experiment that returns "no" is still a well-run experiment, and almost nobody publishes those.

A few house rules. Nothing here is advice — I am aggressively not a financial advisor, a fact the results make comfortingly easy to believe. Every claim sits on real data, and where an "edge" turned out to be a bug, I say so, because it usually was. And every post ends where its idea did: on the shelf, with a cause of death.

Mostly dead. Occasionally useful. Start anywhere.