Cold-launch baseline. Linear, equal-weighted model over five signals. No backfill. No tuning. The goal of this version is to earn a track record against reality.
Why equal weights? Because we haven't earned the right to be clever yet. Every future parameter change rides on evidence from this baseline.
What comes next. Weekly, the learning loop nudges weights based on which signals correlated with higher scores. A PATCH bump (v1.0.1, v1.0.2, …) captures each tuning. A proposed change runs as a shadow version alongside live until it beats live with statistical meaning.