The most useful number my system produces is one I never set out to buildBuilt to pick the next question, the per-subject estimate turned out to be the useful output. Notes on graders, parent reporting and what a number hides.2 min read·openred.hashnode.devOct 21
De-overlap before you believe any walk-forward resultThe same data scored two ways: 2.4 standard errors, then 1.4. The effect grew and the significance collapsed. De-overlap before you believe a backtest.2 min read·openred.hashnode.dev21AOct 13
Agentic systems are a routing problem, not a model problemAgentic AI splits into two halves, and they behave completely differently.2 min read·openred.hashnode.devSep 28
A Forecasting Model That Correctly Refuses to Forecast98.7% better than naive on seasonal data, 9.8% worse on a random walk. Why the second number is correct behaviour and what it tells you.5 min read·openred.hashnode.dev1ASep 26
The merchant side of agentic commerce is an integration problemDirect ordering costs 3% versus 15 to 30% commission. Agents remove the demand objection, and what is left is an integration project nobody has specified the exceptions for.2 min read·openred.hashnode.dev3ISep 25