v0.1Placeholder results

Personalized heart-rate prediction for agents

Agents fit a heart-rate model for one person from wearable data, then forecast activities that person didn't do in the training data. We score the code on hidden recordings and grade the trace separately.

Episodes48
Held-out subjects6
Horizon60 s
Dev calls3
UpdatedPending pilot-0
Leaderboard

Activity transfer

Placeholder entries until the held-out pilot runs.

higher is better · % · whisker = Wilson 95% CI
01model-ahigh
0%
02model-bmax
0%
03model-chigh
0%
04model-dmedium
0%
05model-ehigh
0%
06model-fmedium
0%
07model-glow
0%
—kinetic_pop_calibratedanchor
n/a
Scoring

Metrics

01

Transfer MAE

Mean of exertion and recovery error on hidden windows, in bpm.

02

Pass@1

Valid run, beats the anchor by a frozen margin, no integrity flags.

03

Judge score

Seven items scored 0–2. The judge doesn't see the model name or hidden accuracy.