STRIKE15 · STATE LAB · intra-candle state vs fair value & market · holdout-validated

Reconstructs P(leader settles | state) from tick-level state records and compares it against TWO baselines: a driftless random-walk fair value (is price non-random?) and, where quotes exist, the market's actual price (is the market wrong?). Minimum cell size and split-half holdout are enforced so noise cannot masquerade as edge.

STATE DUMP

Paste a research dump (the states array is what this reads) or a synthetic test file.

Overview — sample & process character

State surface — empirical vs random-walk fair — is price non-random?

State surface — empirical vs market quote — is the market wrong?

Holdout verdict — survives split-half?

Feature ablation — which dimensions actually predict?

Brier score = mean squared error of the prediction (lower is better). Each row adds one dimension to the state key; a dimension earns its place only if it lowers Brier meaningfully out-of-sample. Dimensions that don't reduce error are noise and should be dropped.