Every call is pre-registered. A signal is only scoreable if it was logged before its window opened, with a direction, a probability, a horizon and a threshold. No retrofitting, no "the transit explains the dip" after the fact — that discipline is what was missing from v1.
Probabilities, not prophecies. Pillars emit statements like "62% that gold closes above $4,200 within 5 sessions." These are scored with the Brier score, which rewards both being right and knowing how sure you were.
Weights are earned, not assigned. Each pillar's weight is proportional to its skill above a random-walk baseline over a rolling 26-week window, updated with exponential decay (recent skill counts more). A pillar that can't beat a coin flip decays toward a 5% probationary floor, then to zero after two failed quarters.
The astrology question answers itself. Whether astrology "works" for gold stops being a debate: it is a number in the table above, recomputed weekly. Currently it earns 87% — ahead of the flat-25% baseline every pillar starts at. The engine is agnostic; the scoreboard is not.
Baselines keep everyone honest. The composite is benchmarked weekly against (a) random walk, (b) simple 20-day momentum, (c) "always bullish." If the engine can't beat all three over a rolling year, the whole methodology gets revised — that rule is itself part of the methodology.
Regime awareness. The price-action pillar reads the live weekly/daily trend regime (bull, bear, or neutral) directly from moving-average structure — not a separate "hidden layer," just what the MAs say right now: weekly NEUTRAL, daily BEAR. Tracking pillar skill separately per regime (does an edge hold up differently in trends vs ranges?) is on the roadmap, not built yet.