Holographix: a percept-first codec and network substrate for Large Vision Models and Large Audio Models over UDP

Impaired networks are normal in robotics, remote presence, ad-hoc meshes, radios, and disaster links: loss, burst loss, jitter, reordering, duplication, and intermittent connectivity are the rule rather than an edge case. For perceptual media the relevant objective is not bitwise completeness; it is preserving useful evidence with a reconstruction that improves continuously as surviving fragments arrive. HolographiX implements a field-centric codec and UDP substrate for RGB images and PCM audio. A signal is represented as a coarse scaffold plus an int16 residual: residual = original - coarse_up. The residual is then distributed across a population of chunks by a low-bookkeeping modular interleaver: indices are permuted via a step chosen near N/phi and adjusted to be coprime with N, producing a full-cycle permutation. Chunk loss therefore removes fine detail globally rather than erasing localized regions. Decoding is anytime: missing residual samples are zero-filled and the estimate sharpens as additional chunks arrive. On the transport side, the system targets packet-atomic progress: chunks are sized so that one holographic chunk fits in one UDP datagram under a conservative payload budget, making each surviving datagram a valid contribution without reassembly dependencies. Above UDP framing, the substrate uses content-centric identity (holo://object -> content_id) and a lightweight mesh policy for opportunistic replication, store-carry-forward, and field maintenance (healing) under long-lived loss. The repository includes end-to-end scripts and a reproducible routing impairment testbed (containerlab+FRR+tc/netem) to measure quality-vs-fragments and quality-vs-time metrics.