The two literatures are no longer fully disconnected: 2 edge(s) now join them — Delphi → SafetyNets, Hao et al. → Cheetah. That is worth reading closely; it was zero when this corpus was first built.

They are working the two columns of the same 2×2, on the same operators — GELU, Softmax and LayerNorm are the expensive ones for both — with entirely different hammers, and they are not reading each other.

Counted between the verifiability and privacy columns specifically. A survey citing zkGPT also crosses a section boundary, but says nothing about the two communities talking, so it is not counted here.

No edge runs directly between the two columns. But some papers are cited by both — and those are the real contact surface, more interesting than the absence of a direct edge:

• Cheetah — cited by Hao et al. on the verifiability side, and BOLT, Bootstrapping is All You Need, CipherGPT, Iron, Nimbus, PUMA, Sigma on the privacy side.
• SafetyNets — cited by Hao et al., Proof-of-Learning, Range-Arithmetic, ZEN, ZIP, zkLLM on the verifiability side, and Delphi on the privacy side.
• SecFloat — cited by zkPoT (Garg et al.) on the verifiability side, and BOLT, Sigma on the privacy side.
• SIRNN — cited by Hao et al., zkPoT (Garg et al.) on the verifiability side, and BOLT, Cheetah, CipherGPT, Iron, Nimbus, Sigma on the privacy side.

The pattern is that the contact happens at the numerics layer, not at the systems layer. Both communities are trying to do arithmetic on real numbers in a setting that only offers integers mod p, and both reach for the same small set of fixed-point, floating-point and truncation primitives to do it. They will read the same paper about rounding, and then not read each other's papers about GELU.

See Numerics, which is filed as a cross-cutting section for exactly this reason: it belongs to neither column, because it is underneath both.