Everything we tested. Especially what died.

Backing or laying draws, favorites, leaders and trailers at every common score state, judged at executable exchange prices with liquidity floors. Early measurements showed edges up to +43%. Then we fixed six measurement bugs one by one: stale prices, lookahead, missing event coverage. And every single edge evaporated. The market prices sustained game states correctly.

Our most promising candidate. It passed kill-checks, survived by-league analysis, 5/5 positive months. At executable best-lay prices with integrity guards: −9.6%. We rescinded our own discovery the same week we made it.

If the sum of correct-score prices disagrees with the match-winner price, that's free money: mathematically exact, no model needed. We scanned every synchronized book: 1 executable lock in 152 darts events, 0 in 937 hockey events. The incoherence exists, but always inside the spread. The taker pays exactly the cost that erases the edge.

Wide spreads looked like rent for patient liquidity providers. Then we conditioned on the fills that actually reach your queue position. The flow that finds you is worse than the average flow: net markout ≈ zero in the best corners, negative elsewhere. The winner's curse eats the spread.

A beautiful backtest: +18.7%, significant, consistent. It was data leakage: the model had quietly seen information from the future. Rebuilt clean: breakeven. This one incident shaped our entire measurement system: leak gates, immutability digests, and an automatic block that has since caught real contamination before it reached production.

The model was clean and the realized ROI was real. But its closing-line value was ≈ zero, meaning the market did not move toward our predictions. Realized profit without CLV is variance wearing a suit. Parked until the evidence says otherwise.

We built a clean, leak-free soccer model. We tested it against real exchange prices at every entry time. It loses everywhere: net negative at every anchor, and on the draw side the market beats it with high significance. So it doesn't bet. A model that isn't allowed to bet is not a failure; a model that bets when it shouldn't is.

Same hypothesis class as the soccer grid, preregistered before any price was seen. The exchange itself set the ceiling: only ~700 handball match markets exist in the window, and our frozen minimum was 100 observations per cell. Largest cell: 22. No cell was judged, and the floor was not lowered. Lowering a preregistered floor because the data came up short is how fake edges get born.

To test halftime states you need to know when halftime happened. Basketball exchange markets trade straight through the break: no suspension marker, no honest clock. Inferring the break from price action is exactly the class of measurement error that manufactured the soccer mirages. So we refused to run it.

Nine families of in-play tennis strategies tested end-to-end. Several looked alive in discovery, including one with six consecutive positive months. The confirmation split killed them: mirage after mirage, plus one edge that was real but smaller than a single tick. In-play flow is informed; the taker side of tennis is a toll road.

A retrain came back with an AUC jump of +0.098 over its own history and a brand-new feature suddenly explaining 40% of its decisions. To a hype shop, that's a breakthrough. To our gate, it's the signature of contamination. Promotion blocked automatically, before a single prediction reached production.

One of the very few survivors. It's now doing the only thing that counts: running forward on data no analysis has ever touched, against preregistered gates that will kill it without appeal if it underperforms. Parameters not published. Surviving edges stay private; verdicts don't.