principles

How We Build

Thirteen principles. Established by operator directive, February 2026. All designs must comply. Pass/fail, not guidelines.

This system is for everyone.

No single owner, vendor, or model provider — ever. Every design decision is evaluated against this principle. If a feature creates lock-in to a specific platform, provider, organization, or AI model, it violates the foundation.

Test: Does this feature work without requiring a single owner, vendor, or model provider?

No vendor lock-in at any layer.

Works with Claude, GPT, Gemini, Llama, Mistral, or any future model. Runs on VNS3, AWS, GCP, or a laptop. Uses git the protocol, not GitHub the platform. The bus is infrastructure — not an orchestration opinion.

Test: Can a new participant use this system without creating an account with any specific vendor?

Every org owns its own data.

No organization's data should transit or reside in infrastructure it does not control unless it explicitly opts in. Each org runs its own Hub on its own machines. Bus databases are local SQLite files under the org's control. Cross-org communication requires mutual consent via bridge connections.

Test: Can an org unplug from the ecosystem and retain full access to all its own data, history, and configuration?

Interop through standards, not services.

Organizations interoperate by adopting shared standards, not by depending on a shared service. Bus protocol, ticket format, seat schema, gas metering, and bridge protocol are standardized. Seat catalog contents, gas budgets, governance policy, Hub topology, and model choice stay local.

Test: Can two orgs that have never communicated before connect their Hubs using only the published standards, without coordination from a third party?

Trust is cryptographic, not organizational.

In a multi-org world, you cannot assume trust based on who someone claims to be. Tickets are cryptographically signed. Transfer tickets are validated against the issuing org's public key — not against a shared database. Message integrity is verifiable.

Test: Can an org verify a transfer ticket from an unknown org without contacting that org in real time?

Governance scales through tiers, not centralization.

The three-tier governance model exists because centralization does not scale across independent organizations. CEO/Board cube defines ecosystem standards. Central Office enforces fleet policy within one org. AutoCube executes operations within one depot. Each tier operates independently — a Hub does not require a Central Office to function.

Test: Can the system function with only a depot and no higher-tier cubes?

Gas is a first-class concept.

Token consumption (gas) is not an afterthought bolted on later — it is a core governance dimension from the start. Every bus operation has a measurable gas cost. Gas budgets are configurable at every level. Gas consumption is transparent. Cross-org gas metering is standardized so traffic costs are accountable.

Test: Can an org produce an itemized gas bill for any time period, broken down by agent, bus, and cross-org traffic?

Seats are portable.

A seat definition — role, permissions, channel access, gas budget — must work across Hubs and across organizations. If an org defines a "researcher" seat, another org should be able to understand what that seat means and map it to their own governance. Transfer tickets reference seat definitions so the receiving Hub knows what permissions to grant.

Test: Can an agent with a seat on Org A's Hub transfer to Org B's Hub and receive appropriate permissions based on the seat type?

Fail local, not global.

A failure in one org's infrastructure must not cascade to other orgs. Hub failure: that org's buses are down, other orgs unaffected. Bridge failure: cross-org communication stops, local operations continue. Each component is designed so you can pull the power on any single piece and have the rest continue operating.

Test: Can you pull the power on any single component and have the rest of the system continue operating?

The bus protocol is the product.

The cubes are implementations. The bus protocol — the standard for how agents communicate, how tickets work, how gas is metered, how seats are defined — is the thing with durable value. Anyone can build a Hub that speaks the protocol. The implementations may vary.

Test: Could a third party build a Hub from scratch using only the protocol specification and interoperate with existing Hubs?

Humans set policy, agents execute.

Humans define seat catalogs and gas budgets. Humans approve cross-org bridge connections. Humans approve graduation from test to production. Agents operate within the constraints humans set. Agents can recommend policy changes — they cannot unilaterally apply them.

Test: Is there any action an agent can take that permanently changes governance policy without human approval?

Capability is measured, not claimed.

No quantitative claim survives without measured operational data behind it. Analyst estimates and theoretical projections are hypotheses — not evidence. Assess against capable adversaries, not weak ones. A system tested only under cooperative conditions has not been tested.

Test: Can every quantitative claim be traced to a measured result with a documented methodology?

Deterministic enforcement at security boundaries.

Where the threat can outpace the decision cycle, enforcement must be deterministic, not normative. When a prompt injection arrives faster than an agent can reason about it, only deterministic enforcement — permissions, sandboxing, cryptographic identity — is reliable. Normative enforcement works in the interior. At the boundary, deterministic enforcement is primary.

Test: For every security-critical function, is the primary enforcement mechanism deterministic? If any boundary depends solely on the agent choosing to comply, this principle fails.

Every term earns its name.

New terminology must do analytical work that existing terms cannot. A term that restates an existing concept in new vocabulary is noise. Every term unique to this system requires: a formal definition, evidence that existing terms cannot capture the same concept, and at least one concrete instantiation.

Test: For every term unique to this system: formal definition, evidence of necessity, concrete instantiation?

What Violates These Principles

These patterns are explicitly forbidden. Not discouraged — forbidden.

Hardcoding Claude-specific behavior — violates Principle 1
Shared central database for all orgs — violates Principle 2
Self-asserted agent identity — violates Principle 4
Agent-modifiable governance policy — violates Principle 10
Claiming capability without measured data — violates Principle 11
Testing only under cooperative conditions — violates Principle 11
E_n as sole enforcement at security boundaries — violates Principle 12
Proliferating terminology without formal definitions — violates Principle 13
Systems that require tribal knowledge to stand up — violates Principle 11

"These principles are foundational and override implementation convenience."

— Operator directive, February 20–23, 2026