What does "subsidiarity" mean in this router?

Subsidiarity is the principle that a decision should be handled by the lowest-order authority competent to handle it, escalating only when necessary. The router applies this to fault remediation: it checks the cell-local reflex layer first, then the centralized statistical autonomic layer, then a model-mediated cortex layer, and dispatches to the first one determined to be both authorized and confident - never escalating further than the fault requires.

What happens if none of the automated layers can resolve the fault?

If every evaluated layer - reflex, autonomic, and cortex - declines or abstains, the fault escalates to a human operator. Escalation is the fallback of last resort, not a parallel path evaluated alongside the automated layers.

What is the "cortex" layer specifically?

The cortex layer is where a tiered reasoning model is consulted for faults that have no existing signature at the reflex or autonomic layers. Its proposed remedy is synthesized into a new signature under the registry-clamped authority mechanism described in Claim 58, so a routing decision toward cortex never itself grants the model unbounded execution authority.

Subsidiarity Router for Self-Healing

Executive Summary

AEGIS is not a single remediation engine; it is a small ordered set of layers with different capabilities, latencies, and authorities, and a router that decides, for every diagnosed fault, which layer should handle it. The routing principle is subsidiarity: dispatch to the lowest-order layer that is both authorized to act (per the constitution in Claim 55) and confident enough to act (per the graduation model in Claim 54), and escalate only when that layer declines.

The ordered layers are a cell-local reflex layer for the fastest, narrowest-scope responses; a centralized statistical autonomic layer for patterns that require fleet-wide context; and a model-mediated cortex layer, described in detail in Claim 58, for faults that have no existing signature at either of the first two layers. A human operator is the final fallback, reached only if every automated layer abstains.

Core Innovation: A fixed-order dispatch function that always tries the cheapest, narrowest-scope, fastest-responding layer first, and only escalates - reflex to autonomic to cortex to human - when the layer below has genuinely declined or abstained.

Problem Statement

Most incident-response architectures pick one automation altitude and stick with it: either everything gets handled by lightweight local scripts (fast, but blind to fleet-wide patterns), or everything routes through a central pipeline (fleet-aware, but slower and a bottleneck for high-frequency, narrow-scope faults). Neither approach adapts the altitude of the response to the altitude of the fault.

70%

of outages Uptime Institute's annual survey attributes at least partly to human error or process failure

$9,000

widely-cited (Ponemon Institute) estimate of average cost per minute of unplanned downtime

4

ordered escalation tiers evaluated by the router - reflex, autonomic, cortex, human

Why Existing Solutions Are Insufficient

Local-only agent-based healing (systemd restart policies, sidecar health checks): Fast and narrow, but has no visibility into fleet-wide correlation and cannot escalate intelligently when a fault turns out to be bigger than one node.
Central-only AIOps pipelines: Fleet-aware, but every single fault - however trivial - pays the latency and infrastructure cost of a central round-trip.
Single-tier LLM-agent "autonomous ops" tools: Route everything through a model by default, with no cheaper, faster, non-model layer tried first, and typically no hard authority ceiling on what the model's output is allowed to do.
Simple alert-severity escalation ladders (PagerDuty-style): Escalate based on how loudly an alert is labeled, not based on which layer is actually competent and authorized to resolve the specific fault.

Technical Architecture

The Ordered Layers

Layer	Scope	Latency	Authority Source
Reflex	Cell-local	Fastest	Graduated pairing, Claim 54
Autonomic	Centralized, fleet-aware	Moderate	Graduated pairing + blast-radius constitution, Claim 55
Cortex	Model-mediated, novel faults	Slowest of the automated layers	Registry-clamped signature synthesis, Claim 58
Human	Unbounded	Slowest overall	Full human judgment

Evaluation Order

For each diagnosed fault, the router evaluates the layers strictly in order. A layer is skipped only if it is not authorized for the fault's blast-radius tier or lacks sufficient confidence for the specific fault-remedy pairing; the router does not evaluate multiple layers in parallel and pick the fastest responder. This preserves the lowest-order-first guarantee: a fault that reflex can competently handle never reaches autonomic or cortex.

Escalation on Abstention

Escalation to the next layer - and ultimately to a human - happens only when the current layer explicitly declines or abstains, whether because it lacks a matching signature, lacks sufficient confidence, or is constitutionally excluded from acting on that blast-radius tier.

Subsidiarity Escalation Ladder

The cortex layer is deliberately the last automated tier tried, not the first: it exists specifically to cover faults for which reflex and autonomic have no applicable signature. Because a reasoning model is only ever invoked once the cheaper, deterministic layers have abstained, the system pays the cost and latency of model-mediated reasoning only where it is actually needed. The signature that cortex produces is itself bound to graduation requirements identical to a human-authored signature, so a fault handled by cortex does not receive a lower bar for autonomous trust than one handled by reflex or autonomic.

Dispatch order

Reflex → Autonomic → Cortex → Human

Evaluation mode

Sequential, not parallel-race

Escalation trigger

Explicit decline or abstention

Cortex authority source

See Claim 58

Prior Art Differentiation

System	Layers	Fixed Lowest-First Order?	Model Layer Authority-Bounded?
Local-only agent healing	Single (local)	N/A	N/A
Central-only AIOps pipeline	Single (central)	N/A	N/A
Single-tier LLM-agent ops tools	Single (model)	No	Typically not registry-clamped
PagerDuty-style severity escalation	Human-only ladder	Severity-based, not competence-based	N/A
JIL Sovereign AEGIS	Reflex / Autonomic / Cortex / Human	Yes, strict sequential	Yes, see Claim 58

Patent Claim

Independent Claim 57: A computer-implemented method for routing a diagnosed fault to a remediation authority, comprising: evaluating, in a fixed order, a plurality of candidate remediation layers comprising at least a cell-local reflex layer, a centralized statistical autonomic layer, and a model-mediated cortex layer; determining, for each evaluated layer in order, whether that layer is authorized and confident to resolve the diagnosed fault; dispatching the diagnosed fault to the lowest-order layer determined to be both authorized and confident; and escalating the diagnosed fault to a human operator only upon every evaluated layer having declined or abstained.