i. Cognitive Systems

JUNIPER

Orchestration Layer

The orchestration layer every other instrument runs on: seven departments, local inference first, cloud only when the work genuinely requires it.

Active; the execution substrate for the instrument suite Python Ollama MCP SQLite

What it does

JUNIPER is the orchestration operating system underneath instrument execution. It runs seven internal departments, each with its own tool permissions, and enforces a strict compute hierarchy: work is attempted locally against Ollama-hosted models first, and escalates to a cloud model only when the task genuinely needs capability local inference cannot supply.

That ordering does three things at once. It protects paid quota, it keeps sensitive material on hardware the lab controls, and it keeps the system operable with no network connection at all. An instrument that stops working when the connection drops is not infrastructure.

Departments, not a flat tool list

Departments are the middle tier of the three-tier taxonomy the lab uses: instruments are the named, versioned products; departments are functional agent clusters inside an instrument; task agents are single-function and ephemeral, spawned for one operation and dissolved afterwards.

Muse, one of the seven, carries a protected time protocol that structurally shields scheduled creative blocks from interruption by the other six. That is a systems decision about where creative output comes from, encoded as a permission rather than left to discipline.

In the code

Escalation is a decision the orchestrator makes, not the model
# Local first. A task escalates only when it fails a capability check
# the local model cannot satisfy, never because a cloud model would be
# a little better. The reason is recorded with the escalation so the
# ratio can be audited later.
ROUTE = {
    "local":  {"summarize", "classify", "extract", "score", "condense"},
    "cloud":  {"long_form_argument", "cross_document_synthesis"},
}

Illustrative shape. The important property is that the escalation rule is readable and countable. If the cloud share climbs, that is visible as a number rather than as a bill.

How this differs from the ordinary version

Local is the default path, not the fallback

Most agent frameworks treat a frontier API as the runtime and local models as a degraded mode. Inverting that changes what the system is: the network becomes an optimization, and the instrument keeps its guarantees when the network is gone.

Permissions live at the department boundary

Giving each department its own tool permissions means a capability is scoped to the work that needs it. A department that has no reason to write files does not hold the ability to write files.

In the field

Why a small operation would want this shape

A shop that wants an assistant over its own documents has the same constraint the lab does: the material is sensitive, the budget is finite, and the connection is not guaranteed. Local-first with audited escalation is the arrangement that survives all three.

Questions

What are the three tiers?
Instruments are named, versioned products. Departments are functional agent clusters inside one instrument. Task agents are single-function and ephemeral: they complete one operation and dissolve.
Does anything run in the cloud?
Only work that fails a local capability check. The escalation and its reason are recorded, so the local-to-cloud ratio is a measurable property of the system.