A graveyard with commit history.
Shelfware Labs publishes negative results — trading ideas that were built, tested properly, and didn't work — with the numbers that killed them and the mistakes left in.
Most of what gets written about beating a market is written by people who claim they did. This is the other archive. I built a real prediction-market research stack — a scanner, a paper-trading ledger, an on-chain forensics database — and pointed it at every "obvious" edge on Kalshi and Polymarket: structural arbitrage, sharp-book reference pricing, copy-trading the leaderboard, market-making, weather, cross-exchange spreads. They all failed. This is where the failures are documented, because a well-run experiment that returns "no" is still a result — and almost nobody publishes those.
Shelfware is the industry term for software that gets built and then never used: it sits on a shelf. Everything here is shelfware. The name is not an apology. The code all works; the research is real; the only thing missing is the edge, and its absence is the finding.
How I work
Three rules, and they're the whole method.
Every claim sits on a number. Not "the market felt efficient" — a score for how accurate a forecast was, a measure of skill versus luck, a dollar figure after the fees are actually paid. And every claim has to survive an honest test: I check an idea against real historical data it hasn't already seen, the way you'd only trust a weather forecaster by grading tomorrow's forecast, not yesterday's. Several ideas that looked profitable in a lazy test are reported here as dead, because the honest test killed them.
"Was it a bug?" is asked out loud. An embarrassing share of the "edges" turned out to be mistakes — a probability that summed above one, a Brier score above one (impossible, and therefore a tell), an inverted price field, a units error dressed up as arbitrage. Those bugs are not edited out. They're some of the best material, and catching them is most of the job.
Nothing is smoothed in hindsight. The plan for this whole project was written — by a chatbot, before any code — as a seventeen-hundred-line brief that predicted its own failure in detail. I keep that document and hold each result up against it. It makes the site a pre-registered experiment instead of a story told backwards.
What you won't find here
No signals. No course. No Discord. No "you can still make money if…". If a genuinely exploitable edge existed, publishing it would be a strange thing to do; the fact that this site exists at all should tell you how the search went. Nothing here is financial advice — I am aggressively not a licensed advisor, a fact the results make very easy to believe. Everything is a description of past research on public markets, for people who like watching a hypothesis get taken apart carefully.
Mostly dead. Occasionally useful.