Lab

Instruments & counterfactuals

Descriptive instruments and counterfactual replays. edge_claimed:false throughout -- these surfaces measure, they do not claim.

Counterfactual explorer

Pick a game-state cell -- probability band x game-time -- and see how well the model was calibrated there, next to the market. Reads committed cells; runs no new simulation.

MEASURE calibration error only|edge_claimed: false on this artifact

State-conditioned calibration grid (MLB)

Model calibration error per (band x game-time) cell -- click a cell

mlb in-game

early

mid

late

.8-1

market only

.6-.8

.4-.6

.2-.4

0-.2

market only

well-calibratedhigh errortint = model calibration error

late x .2-.4

source	n	mean p	obs freq	cal err
MODEL	617	0.2620	0.0843	0.1778
MARKET	640	0.2760	0.1594	0.1166

In this cell the model's mean forecast was 0.2620 and outcomes landed at 0.0843; calibration error 0.1778. Lower is better; the market's error here was 0.1166.

Worst buckets = the improvement backlog

Ranked-worst model cells -- shown as a feature, not hidden

mlb / n-weighted ECE

Straight from the artifact story: ranked_worst_buckets is where the model is furthest from outcomes relative to the market -- the improvement backlog. These are the cells to fix next, not numbers to bury.

cell	model n	model cal err	market cal err	read
late((inn7+)) x .2-.4	617	0.1778	0.1166	MODEL TRAILS
mid((inn4-6)) x 0-.2	1,088	0.1379	0.0201	MODEL TRAILS
mid((inn4-6)) x .2-.4	2,285	0.1178	0.1052	MODEL TRAILS
early((inn1-3)) x .2-.4	1,033	0.1162	0.0708	MODEL TRAILS

Aggregate, n-weighted across all mlb cellsmodel ECE 0.0494vs market ECE 0.0397

Read this straight: in aggregate the market is the better-calibrated forecaster here -- its n-weighted ECE (0.0397) is lower than the model's (0.0494). We are not spinning that. The value of this lab is the transparency about which cells the model trails in, not a claim that it wins.

Every figure traces to a committed JSON under scripts/platformkit/analytics_showcase/out/. Grid cells read verbatim from state_conditioned_calibration.json. No ROI, profit, or betting-edge claim appears on this page -- calibration error only.