AION-1 detection channel shifts tomographic mean redshifts by median 0.71 times LSST DESC requirement

AION-1, a 39-modality transformer trained on 200 million objects, was audited by holding image tokens fixed and editing only Legacy Survey segmentation maps. Every derived quantity—flux, size, ellipticity, redshift—changed at rates far above matched controls. The gating mechanism tracks field-centre presence at r=0.47 rather than enclosed light at r=0.30, and ignores pipeline light partitioning across 322 blends with R=-0.006. The Legacy Survey leaves 3.68 percent of targets without segments; propagating these misses shifts mean redshifts by a median 0.71 times the DESC tolerance across 40 bins and exceeds tolerance in 12 bins.

The effect is independent of magnitude weighting and grows with model scale. Withholding the detection channel removes the bias at no measurable accuracy cost, while contradicted photometry degrades performance nine times below the no-metadata baseline. Tokeniser limits compound the issue: image patches resolve only 28 effective states versus 934 for spectra, and the redshift head is quantisation-limited. Sparse dictionary interventions recover 26-75 percent of causal effects and vary 18 points by random seed.

LSST DESC requirements assume unbiased tomographic bins for weak-lensing cosmology. The documented 0.71–8.3 times tolerance violations indicate that current foundation-model pipelines embed survey detection incompleteness directly into cosmological parameter inference. Spectroscopy eliminates the channel; scaling without channel ablation will amplify the systematic.

Next deployments must either excise the detection channel or demonstrate per-bin bias below 0.3 times DESC tolerance before first-light data release.