Skip to content

Cold Start Protocol

Purpose: Describe the procedure for automatically restoring the system from a completely powered-off state to full combat readiness. The protocol covers both a planned cold start (after manual Core Node power-on) and emergency awakening from Dormant Mode or recovery from Spore.

Target time: < 300 seconds until ready to make decisions via Fast Path.


1. Cold Start Stages

Power on → POST/BIOS → OS boot → systemd launches watchdogd → readiness checks → load GlobalState → start isolationd → start vLLM → Decision Pipeline active

1.1. Power On and Hardware Initialization (0–10 s)

  • Motherboard performs POST. External Arduino Watchdog is registered from the start.
  • Watchdog starts timer: if the first valid HMAC-heartbeat from the OS is not received within 120 seconds, Hard Kill occurs.

1.2. OS and System Daemons Boot (10–60 s)

  • Minimal OS boot (read-only rootfs, all non-critical services disabled).
  • systemd starts watchdogd — the first daemon that establishes UART communication with Arduino and begins sending heartbeats immediately after kernel load.

1.3. Readiness Checks (60–120 s)

The readinessd service performs a sequence of checks.

# Check Success Criteria Action on Failure
1 GPU detect All expected GPUs visible in nvidia-smi Retry every 10 s, up to 5 attempts
2 Sandbox test Kata Containers start with GPU passthrough < 500 ms Fallback to gVisor (without GPU), Degraded Mode
3 Network isolation All outgoing traffic except SOCKS5 blocked by nftables Recheck rules, pause
4 TEE attestation Trusted environment attestation valid Transition to DEGRADED, attempt re-attestation
5 GlobalState restore Last snapshot from IPFS loaded and signatures valid Attempt loading previous snapshot

1.4. GlobalState Load (120–150 s)

  • The last valid GlobalState snapshot is loaded from IPFS.
  • Core Node signatures are verified. On mismatch — restore from previous snapshot.
  • ConsistencyRouter applies pending updates from EventBus (if network already available).

1.5. isolationd and vLLM Start (150–250 s)

  • isolationd transitions to INIT → VERIFIED → ARMED.
  • vllm_launcher starts DeepSeek-V4 instances with saved expert masks for Vagrant (20%) and Arbtiragius (30%).
  • Inference check: test prompt → response < 200 ms.

1.6. Decision Pipeline Activation (250–300 s)

  • Decision Pipeline switches to active mode.
  • First cycle: analyze current state, apply urgent security updates.
  • Core Node announces itself in the swarm (if phase ≥ 3) via Nostr Bridge.

2. Cold Start Modes

  • Hard Cold Start – initiated after physical Hard Kill or prolonged power loss. All stages are performed, GlobalState restored from IPFS.
  • Warm Start – initiated after software reboot without power loss. GPU and sandbox state preserved. Readiness Checks abbreviated. Time: < 90 s.
  • Recovery from Spore – special mode initiated by Spore Protocol. Steps: Zombie Seed → MVS search → Core DNA Spore → full state and model load. Expected total recovery time: up to 24 hours.

Black Swan © 2026. Technical preprint. Does not constitute a call to action.