Learned detectors inherit their simulator's noise model: amortized companion detection in Gaia-like epoch astrometry

Gaia DR4 will release epoch-level astrometry in December 2026 and is expected to yield of order 10\textsuperscript{6} astrometric orbital solutions, at a scale where the field's stated obstacle, false-positive orbits, cannot be addressed by human vetting. We present an amortized classifier that maps a source's epoch-astrometry residual sequence to calibrated posterior probabilities over {single star, accelerating, orbital companion} in a single forward pass. On held-out injections the classical χ²-cascade cannot reach 80\% completeness in five of six pre-registered regime cells, so we compare at the cascade's own false-alarm rate: there our method recovers 60.7\% of period-aliasing companions against the cascade's 35.3\%, and recovers approximately 30\% in two regimes where the cascade recovers nothing at all. Per-class expected calibration error is at most 0.004. Under a corrected simulator noise model the advantage narrows to 48.0\% against 36.0\%, due to the detector consuming noise estimates as input features and inheriting a dependency that the cascade's ratio-based statistic normalizes away by construction.