The chain of command ends in code, not in a model.
Every agent in Loriqa operates under a layer of oversight Authorities — coordinated from the Global Management Service down to every running agent. And at the top of that chain sits a deliberate architectural decision: GMS is never under the control of an LLM. An oversight layer that is itself model-driven can be talked into anything. Ours cannot.
Oversight above. Execution below. A boundary between.
The platform is organized into zones. Governance lives in Zone A and watches everything. Agents execute in Zone C. Your governed environment — the systems agents actually touch — sits beyond a trust boundary in Zone D. Nothing crosses that line unobserved.
Agents cannot see, disable, or negotiate with their oversight. The Authorities are not invoked by agents — they are independent, they subscribe to everything, and their decisions are written to the same tamper-evident record as the actions they govern.
The Authorities, in the order your agents meet them
Eleven abbreviations read like a legend. A lifecycle reads like a story. Here is who governs an agent before it exists, while it runs, across the whole fleet, and when something goes wrong.
DPA — the promotion gatekeeper · Deployment Pipeline Authority
No agent reaches production without passing the promotion pipeline — development, test, staging, production — with human approval gates on the way. DPA holds veto authority over every promotion. There is no side door.
FAA — the final examiner · Final Assessment Authority
At the last gate, FAA scores the agent's quality and produces a 0–100 confidence score. Below the threshold, promotion is blocked. Overriding a failed score requires two authorized operators and leaves a permanent record.
CIA — the health monitor · Central Inspection Authority
Validates every agent before it starts, watches it while it runs, and reviews it when it finishes. Agents must actively prove health at defined intervals — a dead man's switch. An agent that goes silent is terminated, not trusted.
RCA — the compliance evaluator · Regulatory Compliance Authority
Evaluates agent actions in real time against the regulations you operate under — GDPR, HIPAA, SOC 2, PCI DSS and more — and produces the evidence bundles your auditors ask for. RCA evaluates the law; it does not make it.
QIA — the model-fit inspector · Quality Insurance Authority
Evaluates whether each step of each task is running on the right model — benchmarking alternatives and scoring them against your priorities: cost, accuracy, speed. A measured answer to a question whose answer changes as the market does.
EPA — the environment sentinel · Environment Protection Authority
Tracks the health of the environments agents run in, so that platform-level degradation is seen as itself — and never mistaken for, or masked by, agent behavior.
AIA — the cross-agent analyst · Audit Intelligence Authority
The only component authorized to read across all agents at once. AIA subscribes to the full event stream and the live heartbeat feed, synthesizes patterns no single-agent view can show, and delivers human-readable findings with the evidence attached.
PDA — the pattern watcher · Pattern Detection Authority
A governed agency of watchers: a supervisor that spawns focused sub-agents, each assigned to a specific data domain, hunting the slow and quiet patterns that never trip a single alarm. What PDA finds feeds the platform's correlation picture.
SRA — the responder · Security Response Authority
The one Authority with the power to act autonomously against a confirmed threat — containing it first and notifying humans while containment is in progress, not before it starts. The time between detection and response is not bounded by human latency. And SRA fails secure: on failure, containment holds. Nothing fails open.
Where the models are allowed — and where they are not
The determinism principle isn't a slogan. It's drawn as a hard line through the architecture, and the MSA family shows exactly where.
MSA — Management Specialist Authority
Four specialist advisors bring LLM-powered analysis to platform governance — and are architecturally prevented from ever acting on it. They are advisory only. They hold read-only access to the event record and cannot write to it. Their models run airgapped and local — never exposed to the internet.
Intelligence where it helps; authority where it cannot be subverted. Analysis may be model-driven. Control never is. GMS — and every decision that binds the platform — remains deterministic code that cannot be prompt-injected, persuaded, or socially engineered.
The abbreviations, in one place
Every term above, defined once more for reference.