Decentralized Mesh Distribution of Skill Artifacts

Mechanism

Decentralized mesh distribution operates by three composed primitives: rateless fragmentation of the skill artifact, lineage-bound activation at the consuming unit, and registry-free authority resolution through the credentialing fabric. The mechanism is structurally invariant under the absence of any single mesh node, any single authoring authority, or any single delivery path. The artifact is not stored at a canonical location; it exists as a distributed cloud of fragments whose reassembly is gated by the consumer's credential set.

In rateless fragmentation, the authoring authority encodes the skill artifact as a stream of erasure-coded fragments using a rateless code such as a Luby-transform or Raptor-class construction. Each fragment is independently signed by the authority and carries a header identifying the artifact, the authority's credential, the rateless degree, and the encoding seed. The fragment stream is potentially infinite; the authority emits fragments at a rate governed by demand. A consuming unit reassembles the artifact from any sufficiently large subset of fragments — typically slightly more than the artifact's information-theoretic entropy — drawn from any source in the mesh. The consumer does not need to acquire fragments from the authority directly; any mesh peer that has cached a fragment can serve it, and the rateless property means peers do not need to coordinate on which fragments to serve.

In lineage-bound activation, the consuming unit reassembles the artifact, verifies the cumulative authority signature against its admitted-credential set, and binds the activation to the lineage of every fragment that contributed to the reassembly. The activation lineage names: the authority, the fragment provenance (which mesh peer served each fragment, which relay path the fragment traversed, which timestamp), the reassembly proof, and the consumer's policy that admitted the authority. Activation succeeds only if the admitted-credential set covers the authority and if no fragment lineage intersects a credential the consumer has revoked. Subsequent invocations of the activated skill carry the activation lineage as part of their own provenance, so any downstream observation produced under the skill is traceable to the artifact, the authority, and the delivery path.

In registry-free authority resolution, consumers discover authoring authorities by traversing the credentialing fabric rather than by querying a central directory. Each authority publishes its credential through the same mesh substrate it uses for artifacts. Consumers admit authorities into their policy by importing the credential through any trusted ingress — peer endorsement, prior credentialed artifact, regulatory injection, or out-of-band attestation. The set of admitted authorities is per-consumer and is not coordinated across the mesh. There is no global authority list; each consumer's authority graph is a local construction.

The mesh substrate carrying the fragments is the spatial-mesh substrate disclosed elsewhere in the parent Cognition Patent. Fragments propagate through fixed-infrastructure relay, peer-to-peer transmission, and mobile store-and-forward. The hop-history relay primitive annotates each fragment with its delivery path, and the consumer's lineage-bound activation consumes the hop history as part of the activation lineage.

Operating Parameters

The rateless fragment size is parameterized to fit the mesh's typical maximum-transmission unit, commonly between 256 bytes and 64 kilobytes. The encoding overhead is governed by the rateless code's overhead factor, typically between 1.05 and 1.20: a consumer must collect roughly five to twenty percent more fragment-bytes than the artifact's raw size to reassemble. Skill artifacts span a wide size range, from a few kilobytes for narrowly scoped tool definitions to hundreds of megabytes for embedded weight deltas, and the rateless construction is invariant under that range.

The fragment-cache policy at each mesh peer is governed by a policy-configured retention budget. Peers cache fragments they have observed in transit, prioritizing recent fragments and fragments whose authoring authority is admitted by the local policy. Cache eviction follows least-recently-used or least-frequently-used heuristics within the retention budget, with the heuristic itself a credentialed configuration.

The activation lineage record sizes in proportion to the fragment count and the hop-history depth. For a typical artifact requiring several hundred fragments, each traversing a few mesh hops, the activation lineage is on the order of tens to hundreds of kilobytes. The lineage is signed by the consuming unit at the moment of activation and is reproducible from the cached fragments and the hop-history annotations.

The authority-admission policy is parameterized by a per-consumer admitted-credential set, a per-credential scope predicate, and an optional revocation feed. The scope predicate constrains which artifact classes a given authority may author for the consumer — a regulator may be admitted only for compliance artifacts, a vendor only for tool-class artifacts, a peer collaborator only for narrow domain artifacts. Revocation feeds are themselves credentialed observations; revocation propagates through the same mesh substrate as artifacts.

Alternative Embodiments

In a pre-staged disconnected embodiment, a consuming unit subscribed to a set of authorities accumulates fragments while connected, reassembles and pre-activates artifacts before disconnect, and operates entirely offline using the pre-staged activations. After reconnect, the consumer ingests revocation feeds, fresh fragments, and updated authority credentials, and reconciles its activation set. Disconnect is treated as a first-class operating mode rather than as a failure case.

In a federated-authority embodiment, multiple authorities co-sign a single artifact. The consumer's admitted-credential policy may require co-signature from a quorum of authorities — for example, a vendor and a regulator — before activation succeeds. The rateless fragments carry the cumulative co-signature, and the reassembly verifies all signatures together.

In a peer-to-peer embodiment without fixed infrastructure, the mesh is composed entirely of mobile peers. Fragments propagate by opportunistic peer contact; no peer is privileged. The activation succeeds whenever a consumer accumulates enough fragments from any peer mixture to reassemble, and the activation lineage records the actual peer mixture that delivered the artifact.

In an air-gapped sovereign embodiment, the mesh is bounded by an administrative perimeter — a national network, a defense enclave, a regulated financial estate. Authorities within the perimeter publish fragments only through perimeter-internal relays. Cross-perimeter import requires explicit credentialed ingress and is logged in the consumer's activation lineage as a perimeter-crossing event.

In a learning-feedback embodiment, the activation lineage records produced at consumers are aggregated by the authoring authority through the same mesh substrate. The aggregate informs the authority's emission policy: fragments for high-utility artifacts are emitted at higher rates, low-utility artifacts are deprecated by the authority through a credentialed deprecation observation. The mesh substrate carries both the artifacts and the feedback in the same provenance fabric.

Composition

Decentralized mesh distribution composes with the spatial-mesh substrate, with the hop-history relay primitive, with the credentialing fabric, with the lineage-bound observation framework, and with the operator-intent envelope. Fragments propagate as ordinary mesh messages. Hop-history annotations contribute to the activation lineage. The credentialing fabric resolves authorities. The observation framework consumes the activation lineage as the provenance of skill-produced outputs. The operator-intent envelope may forbid activation of certain authority classes in certain contexts, with the forbid verdict itself carrying lineage.

The composition is bidirectional. Higher-layer governance primitives consume the activation lineage as evidence about the skill's provenance, while lower-layer mesh primitives consume the consumer's admitted-credential set as policy for fragment caching and forwarding. A peer that does not admit a particular authority may still relay that authority's fragments under a configurable transit policy, since the consumer's admission gate enforces correctness at activation time rather than at transit time. The architecture deliberately separates transit policy from activation policy.

The architecture is invariant under the addition or removal of mesh nodes. Adding a peer adds fragment-cache and relay capacity without coordination. Removing a peer reduces capacity but does not invalidate any artifact, since fragments remain available from other peers. The architecture is also invariant under the addition or removal of authorities; each consumer's admitted-credential set evolves independently.

Prior-Art Distinction

Conventional skill marketplaces — Anthropic's Skills directory, OpenAI's GPTs marketplace, Microsoft's Copilot Studio gallery, Hugging Face's model hub — are operator-mediated. A central platform operator hosts artifacts, gates publication, and serves the canonical copy. Consumers acquire artifacts by direct request to the operator. Operator unavailability invalidates the marketplace; operator policy governs the entire skill economy.

Decentralized package distribution systems — IPFS, BitTorrent, Dat — provide content-addressed retrieval but lack credentialed authority and lineage-bound activation. Federated registries (npm mirrors, container-registry replicas) reduce single-operator dependency but retain the registry abstraction. Code-signing infrastructure provides authority but not substrate-level fragment distribution or lineage-bound activation.

The architecture disclosed here is the structural alternative. Artifacts have no canonical hosting, registries are absent by construction, and activation is bound to the substrate, the credential, and the lineage simultaneously. The disclosure is owned by the parent Cognition Patent.

Disclosure Scope

The disclosure covers any system in which executable agent skill artifacts are distributed as rateless erasure-coded fragments across a credentialed mesh substrate, in which activation at a consuming unit is bound to the lineage of authoring authority and delivery path, and in which authority resolution proceeds without a central registry. Defense, intelligence, sovereign-AI, regulated-finance, healthcare, expeditionary, agricultural, industrial-control, and consumer embodiments are all within scope. The disclosure is independent of the specific rateless code, the specific mesh substrate, and the specific authority semantics.

The disclosure further covers compositions in which the artifact carries executable code, declarative tool descriptions, retrieval indices, fine-tuning weight deltas, prompt-templating libraries, evaluation harnesses, or any other consumable that an LLM-driven agent activates at runtime. The rateless-fragment distribution is invariant under the artifact's internal structure; the lineage-bound activation is invariant under the artifact's runtime semantics. Multi-modal artifacts that combine code, weights, and data within a single signed bundle are within scope, as are compositional artifacts assembled from separately authored sub-artifacts whose individual lineages are merged at activation time.

The architecture's load-bearing claim is the structural absence of a central registry combined with the structural presence of cumulative authority and path lineage at the moment of activation. Marketplaces with replicated mirrors, federated registries with cross-signing, and content-addressed peer-to-peer distribution networks all retain residual centralization at the registry abstraction or lack the lineage-bound activation guarantee. The disclosed architecture removes the registry abstraction entirely and elevates the lineage to a first-class activation precondition. Any system that satisfies both structural properties — no central registry, lineage-bound activation — falls within the disclosure regardless of the surface details of the rateless code, the credentialing fabric, or the consuming agent's runtime stack.