A Field Guide to the Distributed AI Economy

Glossary

45 terms, each tagged by provenance: the guide's convention, kept intact.

W: Defined by the whitepaper
I: Standard industry or regulatory usage
G: Framing introduced by this guide

Accountability loopW: Defined by the whitepaper: The cycle by which validated AI work improves the system that produced it: Frames orient, Cogs perform, Ops orchestrate, Guards validate, Gates control, Tracks preserve, Organizational Memory learns, artifacts improve.
Accountability PlaneW: Defined by the whitepaper: The architecture's name for the validation apparatus as a whole: Guards, Gates, and Tracks together form "not a fourth layer beside the other three, but a plane that cuts across all of them." Revision 9 positions it as this architecture's answer to what the industry calls observability and governance, shipped with the artifacts rather than run beside them. See §6.
AgentI: Standard industry or regulatory usage: Industry term for a model in a loop, with tools, pursuing a goal. This architecture keeps separate three things the word fuses: the capability is a Cog, installable and auditable before it runs; the engagement is an Op, carrying the goal, the autonomy budget, the checkpoints, and the Validation Strategy; and the continuing actor — identity, credentials, memory, and history — is held by the Intelligence Hub. On that reading an agent is a Cog engaged through an Op, given identity and memory by the Hub. See §4.
Autonomy budgetW: Defined by the whitepaper: The boundary between what an agent may do alone and what requires a human. Declared by the Op that carries the engagement and enforced at the Gates that Op defines, which is why the Agent and Gate entries both lay claim to it. Granted per engagement rather than per platform, which is what makes it possible to run a drafting workflow nearly unattended and a payments workflow under approval at every consequential step. See §4, §6.
CogW: Defined by the whitepaper: A discrete AI worker: a model plus a Frame-founded context plus a harness, skills, tools, and APIs plus governance parameters. "An AI worker you can hold to account: not a bare model, but an assembly of model, context, tools, and permissions that can be named, versioned, inspected, and replaced." The atomic, auditable unit of AI work, and the key artifact Nebi distributes in model-heavy, context-heavy, or combined form. See §4.
Context managementW: Defined by the whitepaperI: Standard industry or regulatory usage: The discipline of deciding what a model sees, in what order, under what constraints. In this architecture, Frames form the governed foundation of a Cog's context, with retrieval, memory, and task instructions assembled around them at invocation.
Coordination goodG: Framing introduced by this guide: An economics term for something valuable because others adopt it too (standards, vocabularies, protocols). The guide's explanation for why most Frames circulate free. See §10.
Desktop/Web ApplicationW: Defined by the whitepaper: The application through which knowledge workers run Ops, converse with Frame-oriented Cogs, and install, compose, and share Frames; home of Local Memory. Revision 9 presents it as one worked example of the products around the Hub rather than as the interface layer itself. See §9.
Distributed AI economyW: Defined by the whitepaper: The paper's name for an economy of AI artifacts (Ops, Cogs, Frames, Guards) exchanged across a network of independently owned Intelligence Hubs, rather than through one central platform.
EU AI ActI: Standard industry or regulatory usage: Regulation (EU) 2024/1689, in force August 2024 with obligations phasing in through 2028: enforcement and transparency obligations from August 2026, and high-risk obligations anticipated under the AI Omnibus political agreement for December 2027 (standalone systems) and August 2028 (AI embedded in regulated products). The high-risk provisions require human oversight (Article 14) and automatic event recording with retention (Articles 12, 19), which map onto Gates and Tracks respectively. Noncompliance with the high-risk obligations carries fines up to €15M or 3% of global turnover; the Act's top tier (€35M or 7%) applies to prohibited practices.
FrameW: Defined by the whitepaper: A scoped, versioned, text-based artifact carrying organizational context: rules, terminology, goals, style, norms, skills, tool expectations, prompts, process details, and the Output Guards to be run on results. Scoped, inheritable, composable, shareable, discoverable, and owned: each Frame is accountable to a human or group of humans. Context, in this sense, is capital — a Frame is that context accreting as a versioned asset rather than spent as an expense. See §3.
Frame compositionW: Defined by the whitepaper: Combining multiple Frames for a working session (company + department + project + installed external Frames), with the layering managed automatically.
Frame inheritanceW: Defined by the whitepaper: A child Frame extending a parent (project inherits department inherits company), so context propagates down organizational structure with an auditable chain.
Frame protocolW: Defined by the whitepaper: The open specification for how Frames are structured, making them exchangeable across organizations and implementations.
GateW: Defined by the whitepaper: A decision point in an Op where Guard results determine what happens next: proceed, pause, escalate, retry, or stop. Guards check; Gates decide. Seen from the agent's side rather than the Guard's, a Gate is where an autonomy budget is set. See §6.
GuardW: Defined by the whitepaper: Installable, versioned test-and-verification code that checks whether AI output or action is correct, safe, and policy-compliant. Frames can declare associated Guards; every Op should declare its own. Seven categories in the paper: algorithmic, source-grounding, consensus, expert, policy & safety, regression & drift, outcome. A Guard's market value is held to rest on who stands behind it rather than on its code. See §6.
HarnessW: Defined by the whitepaperI: Standard industry or regulatory usage: The software wrapped around model weights that turns them into something able to act: agent loops, tool-calling frameworks, graph-centric orchestration layers, memory and retrieval scaffolds, skill libraries, evaluation hooks. In this architecture a harness is a capability a Cog builder uses, not a competitor to the Cog; which harness runs inside is meant to be an implementation detail the Cog's consumer never has to know. See §4.
Horizontal and vertical AIW: Defined by the whitepaper: Horizontal AI is an application many industries can use, authored once and installed across many Hubs. Vertical AI is the organization-specific specialization that makes it work against one company's standard operating procedures. The architecture's claim is that a horizontal Op becomes vertical when it picks up an organization's Frames, is checked by its Guards, and leaves Tracks under its governance: "owned context applied to shared capability" in place of custom code. See §10.
Human-in-the-loop checkpointW: Defined by the whitepaperI: Standard industry or regulatory usage: A declared point in an Op where a human approves, refines, or redirects the work before it proceeds.
Intelligence HubW: Defined by the whitepaper: A sovereign, governed AI deployment inside an organization's own infrastructure perimeter: where Frames live, Cogs run, Ops execute, Guards verify, Tracks accumulate under the Hub's governance, and the organization connects to the marketplace. Revision 9 describes it as an organization's Intelligence Infrastructure made concrete, and as where its intelligence and its most important data are brought into intimate contact inside a perimeter it governs. See §7.
Intelligence InfrastructureW: Defined by the whitepaper: The evolution of an organization's information technology into infrastructure that carries its intelligence: models, context, workers, workflows, and evidence, owned and controlled alongside the data they act on. Revision 9's organizing term, introduced in its opening statement and threaded through the executive summary, Layer 1, and the conclusion. The Intelligence Hub is this infrastructure made concrete. See §7.
Intentional vs. emergent contextW: Defined by the whitepaper: Context the organization deliberately encodes (Frames) versus context generated by AI work actually happening (conversations, executions, corrections). Organizational Memory holds both.
Local MemoryW: Defined by the whitepaper: The private, user-side memory in the Desktop/Web Application holding an individual's composed Frames and working context; promotable to shared status only by the user. See §9.
ManifestW: Defined by the whitepaperI: Standard industry or regulatory usage: The declaration packaged with an Op specifying its dependencies, environment, configuration, Frames, and Validation Strategy; what makes an Op installable and auditable.
NebariW: Defined by the whitepaper: OpenTeams' open-source, modular stack for deploying and managing AI infrastructure reproducibly; a curated, additive assembly of open-source components. See §8.
NebiW: Defined by the whitepaper: The packaging and environment-management layer: the common format in which Frames, Cogs, Ops, and Guards are specified, versioned, installed, and rolled back. The architecture's pip/npm analog, and described in Revision 9 as built and working today, on the pixi ecosystem, which was itself built on conda; the project's own site labels it early release, at v0.13 as of this writing. See §8.
NIST AI RMFI: Standard industry or regulatory usage: The U.S. National Institute of Standards and Technology's AI Risk Management Framework (govern, map, measure, manage), with a Generative AI Profile; a declared Validation Strategy is how an Op makes it executable.
Non-rivalryG: Framing introduced by this guide: The economic property that using something does not deplete it. Organizational context is non-rival, which is why sharing a Frame can increase rather than decrease its value to the owner. See §2, §10.
OpW: Defined by the whitepaper: An installable, versioned, supervised program of AI-driven work mapping to a whole job ("close the books"). Composes Cogs, workflow Frames, a supervising model, human checkpoints, integration logic, a declared Validation Strategy, and a manifest. See §5.
Open-weight modelI: Standard industry or regulatory usage: A model whose weights (its trained numerical parameters, the artifact that constitutes the model) are published and can be run, inspected, and hosted by anyone; a precondition for model-level control points inside a Hub.
Organizational contextW: Defined by the whitepaper: The rules, terminology, goals, style, norms, and process knowledge within which an organization's work happens; the asset whose leakage Part I diagnoses.
Organizational MemoryW: Defined by the whitepaper: The Hub's persistent context substrate, accumulating intentional (human-accountable) and emergent context — Guards, Gates, and Tracks included — under the organization's governance, with access controls, retention policy, and the ability to forget. It "belongs to the organization that produced it — owned and controlled," on the reasoning that it is the most intimate part of an organization's Intelligence Infrastructure and that intimacy without ownership is exposure. See §7.
Platform vs. protocol economyG: Framing introduced by this guide: Guide framing: platforms have an owner who intermediates and taxes exchange; protocols have no owner at the exchange layer and win adoption instead. This architecture is structurally a protocol play. See §11.
Products around the HubW: Defined by the whitepaper: The second economy: applications through which people experience a Hub, compute and model management, Track stores, Gate consoles, Op and Cog builders, Guard libraries, segment-specific Ops, and the integration and operations services that keep all of it running. Built by many parties, open and commercial, and integrated into a Hub the organization owns. The Desktop/Web Application is the worked example, not the category. See §9.
Prompt injectionI: Standard industry or regulatory usage: Adversarial input, planted in a prompt or in content a model reads (a document, email, or web page), that overrides the model's intended instructions so it acts on the attacker's behalf. Ranked the top risk in the OWASP Top 10 for LLM Applications (November 2024). See §6, FAQ 11.
ProvenanceW: Defined by the whitepaper: What a marketplace exchanges once generating plausible code is nearly free: verification, context, and accountability with a named party behind them. Tracks that accumulate only through governed use, and Guards backed by identifiable publishers, are the proposed signals. "When generation is free, provenance is the product." See §10.
RAG (retrieval-augmented generation)I: Standard industry or regulatory usage: Supplying a model with retrieved documents at query time. A technique this architecture uses inside Cog context assembly, not a substitute for governed context. See FAQ 3.
Rented intelligenceW: Defined by the whitepaper: The paper's term for consuming AI through vendor-controlled services that the organization cannot inspect, reproduce, audit, or own. See §1.
ReproducibilityW: Defined by the whitepaper: The guarantee that an artifact installed in one Hub is the same system (versions, dependencies, environment, configuration) as in another, "within generative AI limits": environments reproduce; stochastic outputs are verified by Guards instead. See §8, FAQ 13.
ScopeW: Defined by the whitepaper: The defined range a Frame applies to: organization, department, team, project, role, or relationship.
Shared abstractionW: Defined by the whitepaper: An artifact or standard built to be adopted rather than owned, whose value accrues broadly to the communities and companies that adopt it. NumPy, SciPy, conda, and pixi are the worked historical examples; Frames, Cogs, Ops, Guards, Gates, and Tracks are proposed as the AI era's set. The credo term of Revision 9's opening statement and its conclusion. See §8, §13.
Sovereignty (at control points)W: Defined by the whitepaperI: Standard industry or regulatory usage: The realistic form of AI ownership: controlling which model runs, what data reaches it, what policies bind it, and what evidence remains, without owning the full stack. See §7.
TrackW: Defined by the whitepaper: The durable, structured evidence record of an execution: Frames applied, Cogs invoked, Guards run, Gates decided, humans approving, outputs produced. Designed for audit, learning, debugging, and trust; retained under the governance boundary of the Hub that produced it and deliberately not exchanged — "evidence is not for sale." See §6.
Validation StrategyW: Defined by the whitepaper: The declaration, part of every Op's contract, of which Guards run at which stage, where the Gates are, and what the Track retains for how long. See §6.
Vendor lock-inI: Standard industry or regulatory usage: Strategic dependency on a supplier whose interfaces, pricing, and behavior the customer cannot control and whose exit costs grow with use.