Reference

The vocabulary

Movements spread through words people can pick up and use. These are the terms of Cognitive Orchestration, defined for practitioners. The specification of record — the authoritative text — lives at terrene.foundation.

Cognitive Orchestration (CO): The open, domain-agnostic methodology for structuring human–AI collaboration: seven first principles, a five-layer architecture, and a phased workflow with human approval gates. The layer above the execution tools — it governs the collaboration, not the software.
Orchestration (vs. prompting): Designing the institution around the machine, rather than optimizing a single turn (prompting) or a single session (context engineering). You set the boundaries and the standards; the AI executes inside them.
The five-layer architecture: Intent, Context, Guardrails, Instructions, Learning — the stack that holds what an organization knows. Each layer encodes a decision a human makes, which the machine then executes within.
The six-phase workflow: analyze → plan → implement → validate → codify → release. Only two phases (plan, release) require a human gate; the rest converge autonomously. High autonomy, human authority where it counts.
Human-on-the-Loop: The governance posture of CO: not in the loop typing every line, not out of the loop hoping — on the loop, defining the operating envelope and observing outcomes.
Institutional knowledge: What your organization knows — conventions, guardrails, hard-won lessons — written down in a form a machine can read and reuse. In CO it is the asset; the model is the commodity.
The three failure modes: The predictable ways AI fails in production: amnesia (it forgets your instructions), convention drift (it follows the internet's conventions, not yours), and safety blindness (it takes the shortest path, which is rarely the safe one).
Anti-amnesia: Designing so that critical knowledge survives the context boundary — re-injected every session and enforced outside the model — instead of evaporating when the window fills.
Deterministic enforcement: Putting the rules that must never break outside the AI's reach — enforced by code that blocks, not by asking the model nicely. Probabilistic compliance is not enough for production.
The artifact harness: The system that mints institutional knowledge into durable artifacts, circulates them to everywhere you build, observes how they perform, and refines them. The reference implementation is the open WEFT cascade.
The build harness: The targets that consume the artifacts to produce running software — SDKs, project templates, command-line entrypoints, and operational tooling. Where methodology becomes shipped work.
Domain application: CO adapted to a specific kind of work: COC (Codegen), COR (Research), COE (Education), COG (Governance), COF (Finance) — and any vertical you derive. Same principles, domain-specific failure modes.
Built with CO: The mark a team may use — freely and neutrally — to signal that its work is structured with Cognitive Orchestration. Anyone may build with CO; no one owns or controls it.