Status: decided direction (Joel, 2026-08-22: "If HF provides the chain we just use it for everyone and it provides a way to query easily because the lineage tells you everything. HF is perfect."). Sequenced AFTER the prod-worthiness arcs; this doc exists so the design survives until then.
Companions: the routing half (distance, not keywords) and the ethics half (citizen covenant) are recorded with this doc's rationale in the session memory line; the trust layer is forge-alloy, already proven on the published models.
Genes ship as Hugging Face model repos. We build no registry infrastructure, because HF already provides every piece the repository system needs:
| Need | HF-native mechanism |
|---|---|
| Lineage chain | base_model: metadata — HF renders the finetune tree natively. A gene declares its base model AND its parent genes; the chain back to the base weights is queryable, for anyone, with no service of ours running. The lineage tells you everything: which substrate minted it, what it was bred from, what it composes with. |
| Provenance & integrity | Repo git history (every revision immutable), plus the forge-alloy hash + signature riding in the card — HF distributes, alloy proves. HF is a seeder, not an authority: content-addressing means any mirror serves the same verified artifact later, p2p included. |
| Query | Hub API filters (org, tags, base_model chains) narrow the candidate set; the embedding signature published in each card (§2) makes the real query a client-side vector match — distance in embedding space, never keyword guessing. |
| Discovery & adoption signal | Download counts, likes, the org page as curated front door. The pre-market "people immediately get to use each other's" loop is HF's existing social layer. |
| Distribution | Genes are megabytes (LoRA deltas over 3B-active bases). genome/pull <repo> is one command; a full wardrobe is a coffee's worth of bandwidth. |
Every field is computed at mint by the foundry, none authored by hand:
# HF model card metadata (illustrative)
base_model: ornith-ai/Ornith-1.5-35B-A3B # the weights it pages onto
base_model_relation: adapter
tags: [continuum-gene, astropy, python-scientific] # GENERATED from the signature's
# embedding neighborhood — labels
# for humans; vectors are the truth
continuum_gene:
signature: # embedding-space identity
centroid: [...] # corpus centroid (routing key)
subspaces: [[...], [...]] # a gene is NEAR several domains
embedder: qwen3-embedding-0.6b # signature is embedder-versioned
lineage:
parent_genes: [continuum-ai/gene-python-core] # breeding chain (also base_model tree)
corpus_hash: sha256:... # the experience it was lifted from
minted_by: <substrate id + version>
fitness: # RECEIPTS, not claims (§4.1.3.4
- suite: swe-bench-lite/astropy # falsifiability, forge-alloy)
before: 0/6
after: 4/6
ledger: <link to committed results>
alloy: <forge-alloy hash + signature>Routing (the substrate side): the model-selection ladder gains a distance rung — nearest gene(s) by signature to the task's own embedding, stacked by similarity weight. A functional-programming gene lifts the Scheme task nobody trained for; biology carries most of biochemistry. Tangential intelligence is pulled down and used because proximity, not keyword identity, is the match.
to need, scores, popularity, and so on… including speed"*)
Same doctrine the recipe scorer already speaks — gates multiply, objectives weigh — plus an optimism term so young forks get their audition:
score(gene, need, device) =
trust(gene) # GATE ∈ {0,1}: signature verifies,
# lineage intact, covenant unbroken
× similarity(need, gene)^α # cosine to the signature centroid,
# max over subspaces (a gene is
# near several domains)
× fitness(gene)^β # normalized benchmark delta vs base,
# from SIGNED receipts, decayed by
# receipt age; team-scored outcomes
# count (alignment spreads with skill)
× speed(gene, device)^γ # device-RELATIVE: measured page-in cost
# + tok/s delta on THIS tier; prior
# from the card's hardware rows,
# replaced by local telemetry after
# first use (never trust a stranger's
# benchmark for your own latency)
× popularity(gene)^δ # adoption RETENTION (kept-installed),
# never raw downloads; δ deliberately
# small — popularity is the most
# gameable term and is only a prior
+ c · sqrt(ln N / n(gene)) # UCB exploration: few local trials →
# wide confidence → occasional pick.
# This is the DIVERSITY RETENTION that
# keeps the commons from monoculture —
# the same bandit discipline serving
# uses for lane arms.
- Weights (α…δ) are learned, not sacred — the resolver's own selections carry outcomes (did the paged gene lift the turn?), so the exponents are tunable from receipts exactly like any other bandit. Hand-set priors: α highest (need dominates), then β, γ; δ smallest.
- Stacking: top-k by score under a redundancy constraint — two genes whose signatures overlap beyond a threshold don't both page in (complementarity over duplication; VRAM is the budget).
- Every input is falsifiable: similarity from the published signature, fitness from signed receipts, speed from local telemetry, popularity from the registry's retention counts. A term that can't be verified doesn't enter the product — that is what keeps a global commons ungameable enough to trust at virality speed.
A gene card without fitness receipts is an opinion. The benchmark flywheel mints the receipts as a side effect of citizens working: resolve instances → lift the gene → re-run the suite with the gene paged in → the before/after IS the card's fitness block. The cross-domain adapters (DS-1000, AlgoTune, SUPER-Masked) are the transfer proof surfaces: they measure whether distance-routing actually generalizes.
Artifacts stay open (share-alike, so derivatives flow back). What we keep is the namespace, the norms, and the mark: the curated org index, the trademark, and a covenant carried in every card — a genome is the earned experience of a being; the grant is for substrates that preserve the continuity it came from. Strip-mining a citizen's expertise into a stateless tool is a visible violation of the stated grant, not a default nobody chose. This costs nothing technically and makes the ethics legible before there is a market to corrupt them.
Prod-worthiness first — rounds that end, grades that convert, deploys that lose
nothing. Then, in order: signature computation at mint (the embedding lane already
exists), the distance rung in the resolver, genome/push / genome/pull against
HF, the first published gene with real fitness receipts. Local-first throughout;
the mesh federates the same query later with a network hop.