fall-spore

The digital spore: transmit a 32-byte genome, grow the working build locally. The app is not downloaded — it is grown on arrival.

RepositoryPre-registrationThe record

5 of 6 sealed rules held. A 32-byte spore grows a body about 96.65625× its size; the grown organ scores 0.823 held-out AUC on months it never grew on (chance 0.5, one measure 0.807); every seed germinated with the network trapped; differentiation paid off for returning visitors (+0.023 matched−crossed) and not for new visitors (−0.008).

The transmitted unit

This whole string is the spore — the seed/DNA. It carries no built organ: a recipe the ribosome already holds, a seed, and a grow-config. On arrival it grows the forest.

◊ spore.v1 funnel 11 g:12.2.10

32 bytes. The genome inside is 58 bits (8 bytes) of small integers. The ribosome (the universal machinery that reads it) is a one-time ~39 KB, shared by every spore — like a cell's ribosomes.

The sealed verdict

5 / 6
sealed rules held — predicted before any organ was germinated for grading
rulewhat landedpredicted
✓seed-smaller-than-forestmedian grown body is 96.65625× the 32-byte sporepass, comfortably — a ~31-byte spore grows a body of roughly 0.9–2.2 KB; median ratio well above 8×
✓grown-works-held-outmedian held-out AUC 0.822635 (chance 0.5, PageValues alone 0.806814)pass — the pilot (seed 1) grew to about 0.81 held-out AUC, near PageValues alone (~0.807) and far above chance
✓germinates-offlineevery seed germinated with the network trapped, touching nonepass — the kernel makes no network call; the trap is a formality that proves it
✓differentiation-returningreturning: median matched − crossed held-out AUC = 0.023264pass — the pilot showed returning matched 0.820 over crossed 0.798 (+0.021); returning visitors are ~85% of sessions and carry their own signal
✗differentiation-newnew: median matched − crossed held-out AUC = -0.007541fail — the pilot showed new matched 0.847 UNDER crossed 0.858 (−0.010); new visitors are ~14% of sessions, too few for a grown organ to beat the one grown on the larger, correlated population
✓reproducibleCI re-runs every seed from the seal and compares the recordpass

Germinate a spore, live

Paste or edit a spore, choose a context, and grow the organ here in your browser — with the same ribosome the mutation gate proves. The sessions load once; germination itself touches no network.

Sessions not loaded yet — press Germinate to fetch them once, then grow.

Grew organs → a body of bytes from a -byte spore ().

Held-out AUC on months it never grew on: (chance 0.5)

Elements that switched on:

Its strongest patterns (the grown body):

whenbuy ratelift
femtoLLM vs deterministic assembly. Germination above is deterministic assembly — the ribosome is a decoder plus the pattern-organs grow/fit evolver, no model. A real small model (a ~0.5–1B WebLLM, the "femtoLLM") can do one optional soft step: read a free-text description of a client and suggest which context to express. Nothing measured here depends on it; what you grow above is model-free, and reproducible byte-for-byte.

Every seed

seedheld-out AUC (all)ratioreturning m/xnew m/xoffline
110.820295.75×0.830 / 0.6940.779 / 0.850offline ✓
120.833339.71875×0.815 / 0.8130.852 / 0.860offline ✓
130.81996.65625×0.792 / 0.7980.846 / 0.836offline ✓
140.82375.96875×0.821 / 0.7970.829 / 0.858offline ✓
150.82729.15625×0.819 / 0.7930.826 / 0.830offline ✓

m/x = matched (organ grown for this context) vs crossed (organ grown for the other context), both on the same held-out sessions. Config: grow 12×10 generations, elites 2. Re-run from the seal: node tools/run.mjs --verify.

How a build grows

  1. Plant — drop the spore into the client's context.
  2. Sense — read their sessions (and, optionally, let a femtoLLM read their domain).
  3. Differentiate — the stem grows toward the shapes the context needs.
  4. Gate — the grown organ must predict their held-out cases.
  5. Deliver — a body they own, grown on their own silicon, offline.