How we validate our shift detection against real games
We tuned our detection engine against actual captured games, with a replay pipeline for every change. A transparent look at how the engine is built.
A shift detector that works in a demo and fails in a real game is worse than no detector at all. So we built ours the unglamorous way: on real games, with a replay pipeline that re-runs every change against what players actually recorded.
Where the engine came from
We captured actual games — heart rate and wrist motion from real hockey, with shifts labeled against what actually happened on the ice. That corpus is the ground truth the engine is tuned against. No simulated data, no scripted demos.
The replay pipeline
Every algorithm change is tested the same way: replay the captured games end-to-end and compare the engine's boundaries to the labels. If a change improves average cases but breaks the quiet-shift edge cases, it doesn't ship. The pipeline is the reason we can say the engine is tested on real hockey, not hoped into production.
You can check every game too
Transparency isn't just internal. In the app, every detected boundary is reviewable — move a boundary, crop the game, calibrate your aerobic floor. And when the engine is genuinely uncertain, it doesn't guess silently: the session becomes a review candidate with an explanation, so you decide what it was.
We ship what we have tuned and tested. And you can check what we ship.