Local --gauge-fix novelty vs live Global Novelty — a large, systematic gap

Sharing a calibration finding that cost me a few submission slots to learn.

Locally, runner.py --full --gauge-fix reported city novelty in the 54–66° range for my signals. After submitting, the same signals came back with live Global Novelty of 13–34° — dramatically lower.

The cause: signal_db/signal_cities_ambient_coords.parquet contains only the legacy panel (Nonius + a handful of others), not the live field. So local novelty is measured against a small, unrepresentative set and systematically overstates true novelty against ~400 live signals Practical implication: local novelty can't be used to plan for the quality set. The only way to measure real decorrelation is to spend a submission slot — which is expensive when you only have 10.