Concepts

Dependency graph

A continuously updated graph of entities (devices, links, hosts, GPUs, workloads, services, tenants) and the dependencies between them, across domains and vendors. Every other capability queries it. See Cross-Layer Infrastructure.

Resilience

For an endpoint, the number of device-disjoint paths to the network core, capped at two. Resilience 2 survives any single further device failure; resilience 1 does not. Computed for all endpoints in one linear-time pass using biconnected components. See the verification paper.

Compact Contract

A typed, versioned representation of an entity's behavior over a time window that preserves cross-layer relationships and links to raw evidence. Contracts reduce raw telemetry volume by roughly 6,000× and are what Curve-1 reasons over. See Context Compression.

Risk score

A continuous score assigned to each signal by the Curve-1 pre-filter, trained on trajectories that led to failures. It decides where deeper reasoning runs.

Trajectory

A sequence of events over time for one or more related entities. Curve-1 encodes trajectories from every layer into a shared space, so related sequences can be compared across domains.

Hypothesis

A candidate explanation for an incident — an origin and a causal chain — with calibrated confidence and linked evidence. Generated from incomplete evidence; ranked, never presented as certain.

Incident

A group of alarms and signals explained by a common origin. Built with a topology-aware Hawkes process: each alarm is either background or triggered by an earlier alarm on a nearby device, and following the most-likely parent chain leads to the origin. See Causal Incident Intelligence.

Do-no-harm rule

The safety condition every action must satisfy: for each endpoint, resilience after the action is at least the smaller of its resilience before and its required level.

resilience_after ≥ min(resilience_before, required_level)

Churn cap

The maximum number of changes per minute a control plane can absorb before reconvergence causes its own damage. Remediation pacing never exceeds it. See the pacing paper.

Sandbox

An isolated Firecracker microVM in which a single diagnostic or remediation task runs, restored from a clean copy-on-write snapshot before each use.