Notes from the lab. Lessons learned from building ML models, insights from failed experiments, and technical deep-dives into what actually works.
Oct 16, 2026
8 min read
> libero
Scored on an unseen season, a fair value model prices under-21s 19% low and 31 to 33 year olds 16% high, a tilt shuffled folds never show. Carrying last season's age premium forward fixes the veterans and takes the under-21s to 33% low, all of the extra error from reading the premium as a median of round prices. Read as a mean, the lag stops hurting the youngest and still cannot help them.
Oct 3, 2026
10 min read
> karyo
Three mitosis detectors, one unseen tumour type. At the threshold chosen on validation, a 20M YOLO beats a 6M CNN by 0.105 F1 and the CNN cannot be told apart from a ViT. Read each model at its own best threshold and the first gap closes to 0.03 while the second opens to 0.16. Validation had the order right all along.
Aug 10, 2026
13 min read
> var
Multiplying Tempo's 1X2, Blaze's BTTS, and Storm's Over/Under 2.5 into a same-match slip collapses the predicted EV to one identical number for all 12 combinations, and spends 25% of its probability budget on outcomes that cannot happen. Why cross-model slips can't be built by multiplying model outputs.
Jun 23, 2026
14 min read
> var
KairoX Storm evaluated across La Liga, Premier League, and Bundesliga. Bundesliga produced the best log-loss and the only consistently negative ROI, and a Gamma prior added for calibration made every other metric worse.
Jun 9, 2026
12 min read
> var
Seven XGBoost runs on the same La Liga dataset land within 0.0025 log-loss of each other. Their away-betting ROI diverges by 11 percentage points. At that spread, log-loss cannot pick the model to bet with.