methodology
How every number here is earned
The market is efficient. The honest, defensible win is a calibrated predictor that matches the devigged close within noise, plus a measured in-game conditioning improvement. Accuracy is not an edge, and a recorded null is a success. Every figure on this site carries a receipt chip that links back here to define what its verification badge means.
This site reports calibration and sharpness only, never a dollar edge, ROI, or bankroll result. An honest null is a success.
the discipline behind every claim
- Leak-free
- No feature sees anything the market could not see at decision time. As-of joins only; a future-leak fails the gate closed.
- Walk-forward
- Every score is out-of-sample and forward in time: fit on the past, measure on the strictly later fold. No in-sample lift is ever reported.
- Truncation-invariant
- The result must hold when the sample is truncated at either end. A number that only appears at one cut point is a measurement artifact, not a signal.
- Two or more corpora
- A claim must reproduce across at least two independent corpora. A single- fold lift is treated as noise until a second corpus confirms it.
one-command proof harness
Every showcase module ships a self-check. One command runs them all and prints a pass ledger. On the measured run staged into this build, 96/96 modules pass. The number above is read from the staged report at build time, not hand-typed.
python -m scripts.platformkit.analytics_showcase.check_allFresh clone: data-dependent modules fall back to a recorded artifact so the harness stays green on a bare checkout. CI runs the same command on every push; a red run is disclosed, never hidden.
what 'verified' means
Each receipt chip carries one of these badges. Color is meaning: blue is a measured, no-edge result; green is a provisional model win; red is an honest loss shown at full size; amber is pending or stale. Green is never the default.
- edge_claimed:false
- Measured, out-of-sample, and NOT a dollar-edge claim. This is the default and the point: calibration and sharpness only.
- descriptive_only
- A descriptive summary of past data (atlas cards). Not a prediction, not a claim of future performance.
- MODEL_SHARPER_PROVISIONAL
- The model's calibration beat the devigged close on this fold, provisionally. Shown at the same size as a loss.
- MARKET_SHARPER_PROVISIONAL
- The market was sharper than the model here. An honest loss, rendered red, never hidden or softened.
- UNDERPOWERED
- The sample was too small to separate model from market. Neutral, not green: an underpowered result is not a win.
- VALIDATION_PENDING
- The cited artifact was not built on this clone, so the number is not yet measured here. Amber, never silently green.
do-not-claim rail
These figures were measured, then retracted as artifacts. They are listed here so they can never quietly reappear as live results. Each is a documented measurement artifact, cited in docs/JOB_EVIDENCE_PACKET.md, and appears on this site only inside retraction framing.
- +18.38% (a pregame market-follow artifact, not the model)
- 0.119 (an end-of-Q3 win-prob figure with a Q4 leak)
- +54% (an in-play L5-proxy ceiling, not realized performance)
- 78.11 (the same L5-proxy ceiling as an accuracy figure)
- 8.94 / 54.57 (retracted inflated figures)
The full retracted-vs-honest story lives on the retraction page. Nothing on this site claims a dollar edge, ROI, or bankroll result.