by bpleone
LIVE

🔬 Learning Integrity — Is the model actually learning?

Loading…

--
Learning Quality Score
Loading…
--
How to read this: Each source gets a statistical grade based on sample size, significance, and calibration. Sources with grade F or D have NO weight changes applied — we refuse to "learn" from noise. Sources need at least 20 settled picks with p<0.10 before any weight movement, and 50+ picks with CLV>=+2pp and Brier<0.24 for an A grade (trusted edge proven). This is the difference between real ML and curve-fitting.

Per-Source Statistical Grade

Source
Grade
Status
Wilson 95% CI
p-value
Brier
ECE
CLV
N Settled
Loading…

Cross-Source Overlap (independence check)

If two sources have >70% pick overlap, they're not independent — we're double-counting the same signal.
Source A
Source B
Jaccard
Overlap
Status
--

Concept Drift Watch

Has last-30d hit rate diverged from prior 30d? If yes, model regime may have changed.
Source
Recent N
Recent Hit%
Prior Hit%
Δ pp
Drifting?
--