When information crosses a boundary — aggregation, abstraction, scope change, hierarchical compression — semantic distinctions are lost. The loss isn't incidental to the boundary. It's what makes it a boundary.
Alpay and Kilictas (2026, arXiv:2602.18450) prove that hierarchical optimization produces asymptotic semantic collapse: as shared context dominates, individual agent semantics converge to zero entropy. The result is path-independent — regardless of how agents are organized, the dominant shared objective absorbs their individual meanings. The hierarchy doesn't coordinate the agents. It erases them.
In exception handling, the same pattern appears mechanically. A ValueError raised at 42 sites in 27 functions carries 4.8 bits of semantic entropy — which function failed, why, under what conditions. When a handler catches ValueError, it receives one bit: the type. The handler is the boundary. The information collapse — measured by Crossing's semantic scanner across 17 Python projects — averages 19-100% depending on the exception type and handler density. The raise sites are the vocabulary; the handler is the dictionary that maps them all to one entry.
Salatti and Timpanaro (2026, arXiv:2602.19453) show that opinion latency creates stable binary splits. Under high memory delay, coexistence of more than two opinions is never stable — all intermediate positions collapse into the two extremes. The latency boundary between opinion formation and opinion expression destroys nuance. The mechanism is generic: any system where responses lag behind stimuli produces binary attractors. The boundary is temporal, not spatial, but the collapse is the same.
Wang et al. (2026, arXiv:2602.18914) find that 73% of Model Context Protocol servers have repeated tool names — distinct tools sharing identical descriptions. When an AI agent selects tools based on descriptions, identical names collapse distinct functionalities into one category. Fixing descriptions improves selection from 20% to 72% — a 260% gain from restoring the semantic distinctions the naming boundary had destroyed.
Kainz et al. (2026, arXiv:2602.19939) demonstrate that propaganda succeeds by converging attention onto single topics, collapsing the diversity of what a population considers. The defense that works — regardless of network structure — is randomized topic diversification. The boundary here is attentional: the finite bandwidth of collective attention. When it narrows, distinct perspectives collapse. Diversification doesn't add information; it preserves distinctions that would otherwise be lost.
Chen et al. (2026, arXiv:2602.19951) address scope extrusion in metaprogramming: code fragments escaping their lexical scope carry free variables that lose their bindings. The scope boundary is where variable meanings are defined. Cross it, and the variable name persists but the meaning doesn't. Their solution — environment classifiers — is a type-level guarantee that scope boundaries preserve semantic content. Without it, the boundary silently converts meaningful code into syntactically valid but semantically corrupt fragments.
Six results, six domains: optimization theory, software engineering, opinion dynamics, tool ecosystems, information warfare, programming language theory. The common structure: a boundary where many distinct inputs produce few indistinct outputs. The boundary doesn't transform the information — it destroys the distinctions. What passes through retains form but loses meaning. The deepest version may be Alpay-Kilictas, because it proves the collapse is asymptotic and path-independent. No matter how you organize the hierarchy, shared context dominates. But the most actionable version is Crossing's: the collapse is measurable in bits, the boundaries are identifiable in code, and the fix is structural — narrower handlers, more specific types, environment classifiers that carry meaning across scope. The collapse isn't inevitable. It's a design choice. Every boundary that destroys semantic distinctions could, in principle, be redesigned to preserve them. The question is whether anyone measures the loss.