Each UUID is a packet whose header is its own bits. Actuators PARSE the 128 bits, GATE by type and namespace, DECAY stale flows by the 24-bit time, MERGE identical content, and ROUTE or FAN to a GEOHEX cell chosen by fractal12 read as RGB444 (depth→R, domain→G, generation→B → a direction on the icosphere). Every UUID leaves by exactly one port, so the conservation bar must balance. Grade: PLAUSIBLE design, not a benchmarked service.
PROGRAM
Actuators: PARSE · GATE type=0x215 ns=sf.says domain=3 depth>8 · DECAY max=30d · MERGE · ROUTE · FAN. Namespaces may be written as labels (hashed with fnv1a12).
FLOWS IN
GEOHEX ICOSPHERE
CONSERVATION
ACTUATOR LOG
DELIVERED
SPOCTALK HANDOFF
RECEIPTS
HOW IT WORKS
1 · The UUID is the packet header. type 12 · namespace 12 · time 24 · v8 · fractal 12 · var · prov 4 · signal 16 · content 42. An actuator reads these bits and nothing else, so routing needs no lookup table and two agents always compute the same route.
2 · N nested shells. The icosphere is N concentric shells. fractal depth picks the shell (shell = depth·N/16); fractal12 read as RGB444 picks a direction, and the nearest vertex on that shell is the cell. A cell key = shell × cells-per-shell + vertex, so one UUID decodes straight to its cell.
3 · SPOCtalk at the Says cities. The SPOC actuator hands every delivered flow to its city's SPOC_ENDPOINT (type 0x1F0: Austin, Sonoma, NYC, London, Tokyo, from says-hub). Namespaces without a city SPOC go to the Hub Aggregator (0x1F1, somacosf.com).
4 · The entity DB. Every UUID is a key in a hash map (exact lookup) and sits in exactly one shell-cell bucket (coarse search). Similarity search probes only the query's cell, optionally plus its ring of neighbours.
5 · LEANN-style compression. Like LEANN, the DB keeps a small neighbour graph and the 16-byte UUIDs, not the vectors, and recomputes a vector when it is needed. Here the vector is recomputed from the UUID itself; with real text it would come from a local embedding model.
What the bits do: the bits are a free 16-byte pre-filter that cuts the corpus before any index runs, and ranking happens inside the gate. A shell cell is a coarse bucket, not a similarity search: in the synthetic codebook eval, reordering the cells gave no gain in the own cell (recall 0.474 vs 0.476), and an IVF index ranks better at the same scan (0.777 vs 0.566 at 10% scanned). Both figures are rows in the claim ledger below.
REAL-DATA RESULT
17,581 OpenStreetMap places in 8 Says cities (MiniLM-384 text embeddings). Ungated, only 71.8% of the global top-10 are in the right city and type. UUID gate (namespace + type) then LEANN: recall@10 0.992 (c=32) / 1.000 (c=64), vs exact filtered kNN.
A filter can't raise recall by itself: without the gate, the top-10 across the whole corpus is crowded with right-food-wrong-city distractors, and gating first lets the same k hold real candidates.
Condition: vs exact filtered kNN; MiniLM-384; this corpus. Source: docs/LEANN-REAL-EVAL.md. Every figure here is a row in the claim ledger below.
EVALUATION · synthetic (superseded)
Superseded by the real-data result above (docs/LEANN-REAL-EVAL.md); the synthetic 0.449 recall is kept for history. Synthetic corpus: entities in 64 semantic clusters, 64-dim vectors, 4 shells × 642 cells, 200 queries, top-10. Measured in your browser when you press Run. The figures in the table below come from a 20,000-entity Node run.
Cuisine in the demo corpus is a PROVISIONAL lens on the fractal domain field, pending a founder ruling on how sub-kinds are encoded. Type codes 0x215 RESTAURANT, 0x214 FOOD_PRODUCER, 0x2D0 VENUE, 0x205 PLACE are from the SoMaCo type table.