SpiralCore: networking protocols and compression

SpiralCore is a set of geometry ideas (by Matthew) that the language model borrows: a fixed graph of rotations in eight dimensions, with exact bookkeeping of the path taken between two points. This project poses a different question about the same family of ideas: could they help move and shrink data across a network? That question is posed here, not yet tested: the code that exists uses plain zlib framing with no SpiralCore routing. The folder also keeps two SpiralCore references. It is a side project. It is not wired into the model's training or serving, and nothing here is evidence that the model works.

flowchart LR
    A["App or unmodified program"] --> B["Wire layer: UORADP01 framing, zlib or raw per direction"]
    B --> C["Network"]
    C --> D["Wire layer at the far end"]
    D --> E["Service"]
    R1["Reference: SpiralCore v68 HTML"] -. "consulted" .-> M["Model mechanism: SpiralCore/FBS projection"]
    R2["Reference: Atomic Ternary FBS r3.1 PDF"] -. "consulted" .-> M
    R1 -. "future: not built yet" .-> B

Left: the wire transport that exists today. Right: the two preserved references, which were consulted but are not the build source of the model's mechanism (that follows a v63 reference and a hosted v69 file). The dashed line from the reference to the wire layer is a plan, not code.

What is in the project

The project has two parts that arrived at different times.

1. The wire transport exploration (code). One framing for compressed traffic, exposed four ways: a pair of proxies (a plain app talks to a local edge, the edge talks compressed to a gateway, the gateway talks to the service), a socket library the app links directly, a process interposer that makes unmodified programs use the wire (on macOS through DYLD_INSERT_LIBRARIES, on Linux through LD_PRELOAD), and opt-in multiplexing of several streams over one connection (UORADM01). The framing starts with a UORADP01 handshake, then length-prefixed frames with a raw-flag bit, so each direction can send zlib-compressed or raw data. There are 36 tool files (16 Python, 17 shell, 3 C) and a Rust crate, wire-rs, that is not part of the model's Cargo workspace.

This work currently sits at spiralcore-base/wire (82 files). It was copied from a separate review repository at commit 032d903, which keeps the full history. It will move into this folder.

2. The owner-supplied references (documents). Two sources the owner brought in on 13 September 2026, preserved byte for byte:

Source	What it is	SHA-256
`Spiralcore_Dodecahedron_IP_Schema_v68.html` (about 1 MB)	An executable reference: prompt-trace handler, address functions, E8 addressing and operator routing, a Bell-tone codec, breadth-first search, orbit tracing and a Kuramoto phase solver	`2a45c2e5f46c8c36bf7801da1d1e12ee9e2ed8b3da488ff54c20a6ac6c2ceab0`
`atomic-ternary-fbs-reference-r3.1.pdf` (about 3 MB)	"Atomic Ternary FBS Formal Consolidated Reference", revision 3.1	`917cb942811971d63cbe499f3a1641d3360c4dde7ebdc6b87e298c3cf1bbd747`

Also kept: provenance.json (original paths, byte counts and hashes) and a plain-text extract of the mathematics. The scripts in the HTML file were not executed during import, and its embedded fixture claims were not re-run (research/spiralcore-v68/README.md).

How the pieces fit

The HTML reference defines a fixed graph of rotations over the 16 roots of one E8 anchor orbit, and a search that returns an ordered history between two roots. The model's SpiralCore/FBS mechanism follows a related convention (the v63 reference, implemented in crates/uor-r4-core/src/spiralcore_operator.rs, and a hosted v69 file reviewed on 9 October), not this v68 file. The idea is the same family: an output block picks one stored atom and moves it along a short path of signed coordinate swaps (see the mechanism page and the geometry explainer). The wire transport does not use any of this yet. Its framing is plain length-prefixed zlib, chosen for interoperability.

Where it stands

Claim	Kind	Evidence
The wire framing, socket layer, interposer and mux interoperate: 18 of 18 legs of the interop matrix pass	measurement (loopback, one machine; a manual host regression, not CI; needs a built `objpeer`, and the Linux interposer leg needs an mvm microVM)	`spiralcore-base/wire/README.md`, `tools/wire_interop_matrix.sh`; lab record `docs/labs/wire-matrix-repro-2026-10-18/README.md`
1 KiB round trip: wire 41.8 microseconds against plain TCP 55.1 microseconds	measurement (loopback, host-load sensitive)	`spiralcore-base/wire/README.md`
Rust peer, 1.25 MiB compressed: 2.10 / 2.86 / 3.30 ms (min / mean / max); plain TCP 1.52 / 1.92 / 2.16 ms	measurement (loopback)	same
Compression backend `zlib-rs` matches `miniz_oxide` on ratio (0.24489 against 0.24480 on a realistic corpus)	measurement	same
The framing adds no encryption or authentication	proof (by inspection of the design: none is specified)	same; use TLS, WireGuard or SSH outside it
SpiralCore routing helps compression or transport	hypothesis	the README says "no SpiralCore/FBS routing integration"
The references improve the language model	hypothesis	the import record (`research/spiralcore-v68/README.md`) calls them consulted references, not model components

Limits stated by the project itself: loopback numbers depend on machine load; the interposer is macOS and Linux only, numeric IPv4 only, with blocking connect; the residual cost measured is compression CPU, not framing copies or threads.

Where the originals are

The two references were part of the owner's untracked research/ folder. The full original tree is preserved in the private Hugging Face dataset caseyallard/uor-r4-store under research-archive/2026-10-19/tars/spiralcore-v68.tar (3,796,480 bytes, 5 files, SHA-256 dfcb839d896d47d5eda0afe646da72331a03b080acfa55607bf331e9f8091a05). Every file inside is listed in files.sha256 in the parent folder, and the tar is listed in tars.sha256 there. Copies of the originals also remain in the owner's Downloads folder.

Source files brought into this folder

From spiralcore-base/wire: all 82 files (tools, crates/wire-rs, docs, bounded evidence), unchanged.
From research/spiralcore-v68: README.md, provenance.json, the mathematics extract, the v68 HTML (placed under reference/) and the FBS PDF (about 3 MB; stays in the store if the size policy says so, with its hash recorded here).

Go deeper

Records: docs/labs/spiralcore-wire-side-2026-10-18/README.md, docs/labs/wire-counter-aggregation-2026-10-18/README.md, docs/labs/wire-matrix-repro-2026-10-18/README.md, docs/labs/uorobj01-conformance-2026-10-18/README.md.
The older projection records for the model's own use of SpiralCore are in archive/labs/ (see the mechanism page for links).
Related project: wasm-gemm-gnaf/ is another independent side project; it does not use SpiralCore.

UOR-R4