The Seed Library

Knowledge as tiny, readable seeds — not smeared across billion-parameter weights. A universal ribosome grows a working build from each seed on your own data, offline. Carry the whole library in your pocket; plant a seed on any device and it grows the build. ◊ seed.v1

library size — offline-capable 1 real domain 3 realistic-synthetic

Pick a domain to germinate

Each card is a survival-critical domain and the tiny seed that grows its build. The seed carries no weights — only a recipe, a number and a packed genome. Click a card, then grow it offline.

The sealed scoreboard

Per-domain held-out (sealed run)

Imagination (optional, not graded)

On a sovereign node a small local model reads a free-text description of a new domain and proposes a seed; the gate then proves it and the seed stores it. The LLM is the librarian's imagination, not the library. The static page can't run a model, so this is a deterministic stand-in that shapes a candidate grow-seed from your words — it is never part of any graded number.

FAQ

What is a seed?

A tiny konomi tag (~50 bytes) carrying no trained weights — only a recipe pointer, a number and a packed 10-byte genome, plus a signature. A universal ribosome reads it and grows a working build from the node's own local data, offline.

How can knowledge live in seeds instead of weights?

The learned model is an interpretable scorecard — a short list of single-feature threshold rules — which packs into a 10-byte genome. The knowledge lives in a readable, provable, editable seed rather than smeared across billions of parameters. Here the whole four-domain library is 226 bytes.

Does it really work offline?

Yes. Every germination runs inside an offline probe that traps fetch, XMLHttpRequest and WebSocket. The build grows from the seed and local data with no network at all.

Is the data real?

One domain is real (UCI Online Shoppers, 12,330 sessions). The other three are realistic-synthetic, generated from known rules plus noise — an honest stand-in for survival-critical domains, not real-world measurements.

What stops a tampered seed from running?

A SENTINEL-style guard: each seed carries an HMAC-SHA256 signature. The ribosome verifies it before parsing; a tampered or unsigned seed fails the guard and never germinates.

Honest scope. This is a proof on four curated domains — one real (UCI Online Shoppers), three realistic-synthetic (triage, water, crop) generated from known rules — not a full civilizational library. Resilience is architectural: local + tiny + distributed + offline + low-power genuinely survives what kills a data-center; the degree of survival in a real collapse scales with scenario severity and how well the library is curated.

Built on the Konomi architecture, created by Thomas Frumkin. Reuses fall-spore and pattern-organs (grow→fit→grade, the seeded RNG, the stratified split, the AUC, the offline probe) and SENTINEL (verify-before-you-parse). Real data: Online Shoppers Purchasing Intention — Sakar & Kastro, UCI ML Repository (2018), DOI 10.24432/C5F88Q, CC BY 4.0. The page runs the same kernel CI gates (witness mutation gate + tests). Re-run proof & the seal live in the repo.