Current read

Technique votes each state's probability is measured only from bars before it — never from the whole sample

On	Technique	Current state	p(up)	raw	n	95% CI	weight	reliability	share	pull

Reasoning

Cost reality check the constraint that decides whether any of this is tradeable

Backtest walk-forward probabilities, spread charged on every lot change

Position ladder

Rung	edge needed	buy above p(up)	sell below p(up)	lots

Dry run Step Replay 1500 bars Reset

Time	Action	Δlots	lots	p(up)	cost	realised	Why

Technique scoreboard out-of-sample, compared against the instrument's own drift — not against 50/50

Technique	scored bars	accuracy	base rate	lift	Brier	log-loss	skill	reliability

Data quality

Settings saved to config.json; every cached number is recomputed