// Decisive LLC · Partition-Tolerant Edge Synchronization
Continuum keeps distributed nodes converged when the network falls apart. CRDT-based state synchronization for Kubernetes at the tactical edge. No control plane dependency. No manual reconciliation. No data loss.
// The Problem
Traditional orchestration assumes reliable connectivity. Tactical edge environments guarantee the opposite. When the network partitions, the control plane disappears and your nodes become islands.
// The Architecture
Continuum runs beside your existing workloads as four Kubernetes-native components. No external databases. No message queues. No centralized coordination. Every node operates autonomously and converges automatically when connectivity returns.
DaemonSet. The CRDT engine. Handles peer-to-peer state replication and conflict-free merge across all nodes. Persists state to bbolt. Survives pod restarts, OOM kills, and node reboots.
DaemonSet. Tunnel-based cross-cluster networking via WebSocket/HTTPS. Automatic peer discovery with mDNS on local networks and Kubernetes endpoints for in-cluster. mTLS on every link.
Deployment. Cursor-on-Target broker speaking CoT XML on the standard TAK streaming ports (TCP 8087, TLS 8089), plus UDP unicast and multicast SA ingest. Routes CoT across the mesh. Subscription routing by event type and geofence is declared as Kubernetes resources.
Deployment. CRD controller for declarative configuration. All configuration originates at higher echelon and propagates via CRDT. Forward nodes are zero-config.
// CRDT Engine
The sync engine implements five CRDT types optimized for tactical applications. Every type merges deterministically without coordination, consensus protocols, or centralized arbitration.
// Sync Pipeline
State changes propagate through a deterministic pipeline. During partitions, each node accumulates deltas locally. On reconnection, the CRDT merge runs automatically with zero operator intervention.
// Measured
Continuum has been instrumented on real kit over real radios and against emulated bearer profiles. These are the numbers those runs produced, with the conditions they were produced under.
Field — JIFX 26-3, Camp Roberts, May 2026 · 4 nodes, MPU-5 radios, 11 instrumented runs
Bench — DDIL bearer campaign, August 2026 · same kit, emulated bearer profiles
Recovery is bounded by key count rather than by writes made while separated, which is why a 60-minute cut recovers in the same window as a 5-minute cut. August figures are bench emulation on the same four-node kit, not radio field data; field re-validation of the post-August stack has not yet been run. Contested-HF delivery is partial while the link is impaired and drains to zero loss on recovery. Continuum is at TRL 6 with no production deployment to date.
// Deployment Architecture
Continuum runs on anything that runs K3s. From a cloud-hosted TOC to a Beelink mini-PC in a rucksack. Same container image. Same CRDT engine. Same mTLS mesh. The only difference is what you plug it into.
Nodes enroll with a bootstrap token and a peer address, then pull their certificates automatically. Configuration originates at higher echelon and propagates via CRDT, so forward nodes need no local admin and no configuration management.
When backhaul drops, nodes on the local mesh keep sharing state. When the link returns, upstream catches up automatically. No phone calls. No manual re-sync. No lost tracks.
TAK clients and CoT-emitting sensors connect straight to Continuum's broker on the ports they already use. It adds partition tolerance and persistent state to TAK's real-time SA rather than replacing any of it.
// Security Architecture
Mutual TLS on by default across every peer, mesh, and federation connection, with handshake identity bound to the verified peer certificate. No relaxed security for forward nodes. The same policy enforcement at the far edge as at the TOC.
FIPS-approved cipher suites are enforced at the application layer on every TLS constructor, and a BoringCrypto build is available. No CMVP certificate has been obtained and no claim of FIPS validation is made.
// Why Continuum
Continuum doesn't replace K3s. It runs alongside it. K3s handles local container scheduling. Continuum handles making sure every node eventually knows what every other node did, even when they can't talk to each other.
TAK provides real-time situational awareness. Continuum adds what TAK cannot do alone: state that persists across a partition, writes that queue locally and replay in priority order on reconnect, and peer-to-peer reconciliation with no server in the path.
No centralized control plane. No master node. No consensus leader. Every node in the mesh can operate independently and contribute to the collective state. Lose any node. Lose any link. The mesh adapts.
CRDTs aren't eventual consistency by hope. They're eventual consistency by proof. The merge function is commutative, associative, and idempotent. Regardless of message order or delivery timing, all nodes converge to the same state.
Built for platoon-to-company scale, characterized at four nodes in the field and ten on the bench, with hierarchical federation above that. Delta-encoded state over constrained links. Not a data center product forced into a rucksack.
Every state carries a sync priority from 0 to 7, and deltas transmit in priority order. Blue-force tracks and C2 tasking preempt routine traffic on a constrained bearer. Measured under twice the link's sustainable load: 96.7% on-time at priority 7 against 67% for bulk.
Each source writes only what it asserts. Position, callsign, classification, and staleness fold independently per entity, so a position-only sensor cannot blank a callsign it never carried. Provenance survives the merge, and genuine disagreement between sources stays representable instead of being silently resolved.
Peer sync runs over QUIC or gRPC with automatic per-peer selection and fallback. QUIC is what makes high-loss bearers usable: TCP goodput collapses under sustained loss where QUIC does not. Ack-clocked pacing keeps a constrained link from being handed more than it can carry.
Standard CRDs. Helm charts. DaemonSets. Deployments. No custom runtime. No proprietary abstractions. If you know K8s, you know how to deploy and operate Continuum. Go binary. Linux only. No bloat.
// Get In Touch
Whether you're evaluating edge orchestration for DDIL environments, looking to add partition tolerance to your TAK deployment, or want to see a live demo of CRDT-based state convergence, let's talk.