1 · The UUID math
A GYST UUID is 128 bits split into fields. Change any input and watch the bits, the hex and the routing address change.
2 · GeoHex: the icosphere
Start with an icosahedron (12 vertices) and split every triangle into 4, pushing new points out to the sphere. Level n has Cn = 10·4ⁿ + 2 cells, and exactly 12 pentagons at every level (the original 12 vertices); all other cells are hexagons. N of these spheres nested inside each other are the shells.
3 · The Says Network + SPOCtalk
The Says Network is a set of city intelligence sites (citysays.com). Each city's domain hashes to a 12-bit namespace, and each city runs a SPOCtalk agent, a SPOC_ENDPOINT (type 0x1F0). The hub at somacosf.com is the SPOC_AGGREGATOR (0x1F1). A flow whose namespace matches a city goes to that city's SPOC; anything else goes to the hub.
4 · RustGIS in Imbue (imbue-geo)
Imbue's imbue-geo crate pins opengeos/RustGIS (Apache-2.0, rev a5fa7c4, engine layers L0–L3 only) behind Imbue's capability gates. It gives native Rust projection and geometry, for example Web Mercator:
GeoHex uses the same lon/lat a different way: lon/lat → a unit vector on the sphere → the nearest icosphere cell. So a place has two lenses: a geographic cell (where it is) and a semantic cell (fractal12 as RGB444: what it is). The city table shows both.
5 · How we harvest LEANN
LEANN (MIT, MLsys 2026 best paper) is a local vector DB that saves ~97% storage. Following the harvest rule (take the capability, never clone the UI), here's what we keep and what we leave:
| Take | How UUID Flows uses it |
|---|---|
| Graph-based selective recomputation | Store a small neighbour graph and the 16-byte UUIDs, not the vectors, and recompute a vector only when a search touches it. |
| High-degree-preserving pruning, CSR graph | One compact graph per shell-cell bucket; the cell already did the coarse cut. |
| Local, private, MCP-ready | Runs beside Imbue as a sidecar or MCP tool, so no data leaves the PC. |
| Leave | Why |
| The Python app, LLM chat and data-source integrations | Imbue already has its own panels and connectors. |
Measured in our harness: 82% storage saved, a ~62× faster probe than brute force, recall@10 0.45 on a synthetic corpus. The bits are a pre-filter, not a similarity search: ranking happens inside the gate. Superseded by the real-data result: on 17,581 OSM places the UUID gate + LEANN reaches recall@10 0.992 (c=32) / 1.000 (c=64) vs exact filtered kNN (docs/LEANN-REAL-EVAL.md). Real text embeddings via LEANN's recompute path: now measured in the real-data eval linked above.