With ten global parameters (zero per-planet dials) solved on a training split and confirmed on a hash-frozen held-out split, the model beats a strong geometric-regularity null on real exoplanet systems — and wins precisely where genuine non-geometric structure exists.
The model’s distinctive prediction — which systems host giant planets — carries no independent information beyond stellar metallicity and mass. Its giant signal is a noisier proxy for the star it was built from.
A CLI-driven forward model that computes planet positions from disk physics, replacing v1’s 30 per-planet dials with ten literature-motivated global parameters. Boundaries (dead-zone edges, snow line) are derived; positions are outputs; every position-determining event emits a structured mechanism-log record. Validation ships null models and Monte-Carlo chance rates computed before scoring, so the instrument can — and does — fail.
The project’s honest verdict, in one line: the physics captures the deterministic geometric organization of planetary systems (Hill-regulated “peas-in-a-pod” spacing) near the achievable limit, but adds no independent information about planet demographics. It earns a real, falsifiable spacing leg — not a demographics one.
The original v1 browser tools, kept for reference. Their headline claims (99.6% accuracy, “validated on 43 planets”) were retired after the v2 audit — they came from per-planet tuning and a validation test measured at 100% chance saturation. Explore them as exploratory instruments, not results.