Coverage ≠ Exposure: Auditable Control of Same-Support Tail Failures under Multimodal Missingness

Multimodal systems encounter partial observability from sensor dropout and degradation, yet adverse tails can remain unreliable even when average performance is preserved. Under a locked same-support contract, matched-average runs can exhibit same-support tail failure, diverging on worst-case and shift-reweighted metrics over the same observable support. The failure arises because conditional interaction separates environment coverage from parameter exposure. Gated routing sends information through different groups, so high-leverage groups may receive few updates despite complete coverage, and tail aggregation amplifies their errors. We audit this mismatch with TailPressure, an exposure-normalized statistic of tail-leveraged interaction from gating logs. We propose Heterogeneity-aware Closed-loop Exposure Stabilizer (H-CES), a lightweight controller that stabilizes per-group pressure via deterministic increment-branch gating and group-wise decoupled weight decay, without changing loss or inference. Across multimodal settings and backbones, H-CES improves same-support tail reliability while preserving clean performance.

Modern AI systems often combine information from multiple sources, such as images, text, audio, or sensors. In real-world settings, some of these inputs may be missing, noisy, or degraded, especially in rare but important situations. A system can look reliable on average while still failing badly in these difficult cases, which makes it hard to trust in practice. This paper studies why such failures happen in multimodal AI models. We show that seeing a wide range of input conditions during training does not necessarily mean that every part of the model learns equally well. Some parts may be used only rarely, and their errors can dominate performance when the system faces challenging missing-input cases. To make this problem easier to inspect, we introduce an audit measure called TailPressure, which helps identify when important parts of a model are undertrained. We then propose H-CES, a training method that keeps model components more evenly exposed during learning. Across several multimodal tasks and model types, H-CES improves reliability in difficult missing-input settings while keeping normal performance stable.